Weakly supervised classification method based on graph disambiguation correction network

Abstract

This invention discloses a weakly supervised classification method based on a graph disambiguation correction network. This method is applicable to the classification of partially labeled data with ambiguous labels in the label space. Specifically, it includes the following steps: (1) The user selects the samples to be classified from the sample (including images, documents, audio, etc.) storage device; (2) Extract the corresponding features according to the characteristics and features of the samples themselves, and organize the ambiguous label set of the corresponding samples; (3) In this invention, we propose a graph disambiguation correction network. The algorithm consists of two parts. The purpose of the disambiguation model is to disambiguate the labels of the candidate label set so that they have real labels, in order to "deceive" the correction model; the purpose of the correction model is to distinguish the correctly disambiguated instances from the incorrectly disambiguated instances as much as possible. Specifically, for a given instance, the disambiguation model tries to fit its potential real label by the label distribution on the candidate label set and its connected instances, in order to deceive the correction model, while the correction model tries to detect whether the disambiguated label is accurate; (4) The correction model detects instances with incorrect disambiguation, thereby adjusting the weight of the candidate label of the instance, reducing the influence of the incorrect label, and re-disambiguating the label. In this way, the two can alternately and iteratively improve the performance of the model. Furthermore, when considering the implementation of the disambiguation model, a novel method for calculating the partial label cross-entropy loss is proposed to overcome the limitation that the traditional cross-entropy formula cannot be applied to ambiguous labels. When considering the implementation of label correction, a novel label probability threshold method is proposed to correct the labels that have been incorrectly disambiguated; (5) If the user is satisfied with the prediction result of the previous step, the process ends; otherwise, more samples are selected from the sample storage device for training, and the process proceeds to step (2). Most existing partial label algorithms directly perform disambiguation on candidate labels, but lack the correction of the disambiguation result. This makes the partial label algorithm susceptible to error-prone samples. In contrast, this invention, based on the graph disambiguation correction algorithm, can make full use of the potential relationship between instances and candidate labels, and continuously correct the disambiguated labels through the idea of ​​game theory, thereby improving the accuracy and generalization performance of the algorithm.

Description

Technical Field

[0001] This invention relates to the field of computer application technology, specifically to the phenomenon of ambiguity in sample labeling, and belongs to the field of weakly supervised classification technology. Background Technology

[0002] In practical applications of machine learning, classification is a fundamental subtask for many problems. In the internet age, acquiring digital samples has become increasingly easy, and massive amounts of information such as images, documents, and products exist across various industries. Efficiently and accurately returning the user's search target, or performing precise retrieval and recommendations, all rely on the efficient classification of sample information. However, with the development of the internet, the amount of data has increased dramatically, and the amount of data requiring labeling is also growing rapidly. Labeling data itself is very time-consuming; therefore, how to automate the labeling of massive amounts of data is currently a hot research topic in machine learning. This type of method treats the classification process as a learning process. First, it selects some samples with ambiguous labels, submits the extracted features and corresponding ambiguous labels to a partial label learning algorithm to learn a classifier, and then sends the sample to be classified to the classifier to obtain the category label for that sample.

[0003] Most existing machine learning algorithms are based on the assumption of strongly supervised learning, which requires a large number of clearly labeled training samples to obtain a classifier with strong generalization performance. In today's big data era, obtaining a large number of training samples is not difficult. However, due to the characteristics and costs of the samples, labeling each sample individually requires a significant amount of manpower. Furthermore, it is difficult to ensure that the number of samples of each category is equal when acquiring a large number of samples. For example, a news image may contain several faces, and the news narrative may include several related names. Mapping each face to a name would require a large amount of manpower. Directly assigning these names as ambiguity markers to the face image samples is less efficient. Using the learning algorithm of this invention for training and classifying the test samples automatically completes the labeling. However, current classification algorithms are all based on accurate labels and cannot classify samples with ambiguous labels. Most existing partial label learning algorithms disambiguate candidate labels based on the similarity of instance features, which ignores the potential information in the candidate label set and lacks correction of the disambiguation results. This makes partial label algorithms susceptible to error-prone labels. In contrast, this invention, based on a graph disambiguation correction algorithm, can make full use of the potential relationship between instances and candidate labels and continuously correct the disambiguated labels through the idea of ​​game theory, thereby improving the accuracy and generalization performance of the algorithm. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to solve the problem that existing classification techniques rely on a large number of clearly labeled samples and insufficient utilization of candidate label information. To this end, a weakly supervised classification method based on graph disambiguation correction network is proposed.

[0005] Technical Solution: To solve the above problems, the present invention provides the following technical solution: A weakly supervised classification method based on graph disambiguation correction networks, which specifically includes the following steps: (1) Construct a set of ambiguous labels according to the pre-established rules and select samples from the sample storage device; (2) Extract corresponding features based on the characteristics and features of the samples themselves, and organize the ambiguous label set of the corresponding samples; (3) In this invention, we propose a graph disambiguation correction network. The algorithm consists of two parts: the disambiguation model aims to disambiguate the labels of the candidate label set to make them have true labels, thus "deceiving" the correction model; the correction model aims to distinguish correctly disambiguated instances from incorrectly disambiguated instances as much as possible. Specifically, for a given instance, the disambiguation model attempts to fit its potential true label by the label distribution of the candidate label set and its connected instances, thus deceiving the correction model, while the correction model attempts to detect whether the disambiguated label is accurate. (4) By correcting instances of incorrect disambiguation detected by the model, the weights of the candidate labels for those instances are adjusted to reduce the impact of incorrect labels, and label disambiguation is performed again. In this way, the two processes can be alternately and iteratively improved to enhance the model's performance. Furthermore, when considering the implementation of the disambiguation model, a novel method for calculating the partial label cross-entropy loss is proposed to overcome the limitation that the traditional cross-entropy formula cannot be applied to ambiguous labels. When considering the implementation of label correction, a novel method for setting a label probability threshold is proposed to correct incorrectly disambiguated labels. (5) If the user is satisfied with the prediction result of the previous step, the process ends; otherwise, more samples are selected from the sample storage device for training, and the process proceeds to step (2). Furthermore, in step (2), corresponding features are extracted based on the characteristics and features of the samples themselves, and the ambiguous label set of the corresponding samples is organized. Specifically, the sample set is defined as follows: ,in yes d 3D eigenvectors Representation and Sample The corresponding set of ambiguous labels, the true label of the sample .in Represents the output space Class tags. Given a candidate label matrix, if the instance Belongs to the tag ( If the membership relationship is non-negative, then the non-negative strength of the membership relationship is... ,otherwise, However, during the training process, Authentic Labels In its ambiguous tag set (Right now In this context, learning algorithms cannot directly access the true values ​​of instances. To address this issue, it's necessary to fully utilize the latent information between instances and candidate labels. Therefore, adjacency graphs are constructed based on the correlation between instance features and candidate labels. The adjacency graph for instance features... If there is an edge between two instances, then ,otherwise The adjacency graph of the label features ,Depend on It is calculated, among which and Examples and The non-negative membership vector.

[0006] Further, in step (3), the purpose of the disambiguation model is to disambiguate the labels of the candidate label set so that they have true labels, thereby "deceiving" the correction model; the purpose of the correction model is to distinguish correctly disambiguated instances from incorrectly disambiguated instances as much as possible. Specifically, for a given instance, the disambiguation model attempts to fit its potential true label by the distribution of labels on the candidate label set and its connected instances, thereby deceiving the correction model, while the correction model attempts to detect whether the disambiguated label is accurate, specifically including: (1) Based on equation (1), the function can be maximized and minimized alternately. To learn the optimal parameters for the disambiguation model and the correction model. In each iteration, the disambiguation labels and non-candidate labels obtained through the disambiguation model are... To train the correction model Among them, the disambiguation model In the correction model Guided by the algorithm, the parameters are updated to approximate the true label distribution from the candidate label set. Iterative competition between the two models drives each other to improve performance until the model output no longer changes.

[0007] Furthermore, in step (4), by correcting instances of incorrect disambiguation detected by the model, the weights of the candidate labels for those instances are adjusted to reduce the impact of incorrect labels, and label disambiguation is performed again. This allows for iterative improvement of model performance through alternating iterations. In addition, when considering the implementation of the disambiguation model, a novel method for calculating the partial label cross-entropy loss is proposed to overcome the limitation that the traditional cross-entropy formula cannot be applied to ambiguous labels. When considering the implementation of label correction, a novel method for setting label probability thresholds is proposed to correct incorrectly disambiguated labels, specifically including: Disambiguation Model: To find the correct label for each training instance, we designed a novel bi-graph convolutional autoencoder partial label learning algorithm. This model has the following capabilities: 1) using the instance adjacency matrix... Adjacency matrix of candidate labels 1) Map instances and candidate label sets to a low-dimensional embedding space; 2) Dynamically update the confidence relationship between instances and candidate labels while learning new representations of instances and label nodes; 3) Achieve ambiguity resolution between instances and candidate label sets by reconstructing the ambiguity of our initialized labels.

[0008] Bi-Graph Convolutional Encoder: Graph convolutional layers contain neighbor information, and the latent information of feature values ​​is passed and transformed between graph edges. We leverage this property by using instance adjacency graphs during propagation. Adjacency graph with labels To obtain more representative embedding representations, we aim to predict link weights by continuously learning node representations from uncertain link relationships. Therefore, dynamically adjusting the propagation weights between nodes and their neighbors by considering the features themselves is essential. To learn rich graph patterns introduced by different edge types, we employ a two-layer Graph Attention Network (GAT) as the message-passing and encoder layers of a dual-graph convolutional encoder, allowing... For the first The output of each convolutional layer Its form is as follows: (2) in It can be an instance or a tag (abbreviated as) or ), and It consists of the instance adjacency matrix and the label adjacency matrix. and . This indicates a multi-head attention mechanism. For learning parameters, The attention coefficient can be calculated using the following formula: (3) To achieve the final embedding of instance nodes and candidate tag nodes, we convert the intermediate output instance encoding and candidate tag encoding as follows: (4) Bilinear Decoder: For the disambiguation candidate label link weights in the instance graph, we consider using a bilinear decoder, which performs bilinear operations and then applies them. softmax The function generates a probability distribution over possible candidate labels: (5) in For training parameters, the model employs a non-linear transformation that converts the representations of the instance confidence matrix and the candidate label confidence matrix into a unified embedding representation. After connection, the confidence matrix is ​​passed through a fully connected layer network to obtain the final label confidence matrix, calculated as follows: (6) To optimize the disambiguation model, we minimize the partial label cross-loss function while using a graph Laplacian regularization term in the loss function to smooth out candidate label information. (7) (9) To optimize the disambiguation model, we minimize the partial label cross-loss function while using a graph Laplacian regularization term in the loss function to smooth out candidate label information. (10) in The proposed method for calculating the partial label cross-entropy loss updates the dissimilarity matrix through iterative calculation. This allows the model to converge optimally to the potential true labels, thereby achieving the purpose of disambiguation. The purpose of this is to smooth the label information so that similar instances in the feature space can share similar labels.

[0009] Correction Model: In contrast to the disambiguation model, the discriminative model aims to maximize the log probabilities of the disambiguated label and the non-candidate label. Traditional graph network computation requires the entire graph as input, which has the following drawbacks: 1) The computational cost increases exponentially with increasing data; 2) The trained model is difficult to apply to newly added data points. Therefore, it is crucial to learn methods to generate effective node embeddings for unknown data using node feature information. This invention overcomes the shortcomings of traditional graph networks by continuously randomly sampling subgraphs from a large graph and approximating the large graph through unbiased estimation. Through graph sampling, a complete discriminative model is constructed to generate embeddings and calculate the loss for each embedding, which can be calculated as follows: (8) in Yes, the disambiguation label of the subgraph node. Yes, the membership matrix corresponding to the complement of the candidate labels of the subgraph nodes. This corrects the label confidence score of the model output. The probability threshold for the label is calculated using the label confidence score matrix of the corrected model output. (9) in Then it can be done Correct any potentially misdisambiguated labels, and update the initial confidence function by correcting the model's predictions as follows: (10) in It is a diagonal matrix. Therefore, the accuracy and generalization performance of the model can be continuously improved by combining the disambiguation labels of the disambiguation model with the label correction of the correction model.

[0010] Beneficial effects: Compared with the prior art, the present invention has the following advantages: Existing machine learning classification techniques all require a large number of accurate labels. This invention can use samples with ambiguous labels for training and predict the test samples, which can greatly reduce the cost of manual labeling, avoid errors in manual labeling, make full use of the potential relationship between instances and candidate labels, and continuously correct the disambiguation labels through the idea of ​​game theory, thereby improving the accuracy and generalization performance of the algorithm. Attached Figure Description

[0011] Figure 1 This is a flowchart of the weakly supervised classification method using graph disambiguation correction networks; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a flowchart of steps 3 and 4 of the method of the present invention. Detailed Implementation

[0012] The invention will be further illustrated below with reference to specific embodiments.

[0013] like Figure 1 As shown, the sample (including images, documents, audio, etc.) storage device contains all samples to be classified. First, a subset of samples is selected for training. For the original samples, feature extraction is required. In this step, an appropriate method must be selected based on the characteristics of the sample itself. For example, for images, convolutional neural networks or principal component analysis can be used to extract sample features; for document samples, keywords can be selected as features, and word frequency, word frequency-inverse document word frequency, etc., can be used to construct dimensionality-reduced features; for audio, audio attribute feature vectors can be generated. It is worth noting that in practical applications, after constructing the annotation set, there is often a situation where the number of samples is too small, which will reduce the generalization performance of the classifier. In this invention, it is stipulated that there is only one true label for a sample. For example, for naming faces, one face corresponds to only one name; for a document, only the most accurate category is defined. After obtaining the sample features and ambiguous annotation set, the method of this invention is used to train the set and classify the test samples, such as... Figure 1 As shown, if the sampling accuracy meets the standard, the process ends; otherwise, the model performance is improved by using user feedback and increasing the sample size to obtain a new ambiguous sample classification model, which is then retested.

[0014] The method involved in this invention is as follows: Figure 2 As shown. Step 10 is the initial action. First, features are extracted based on the characteristics of the samples themselves, and the corresponding candidate label set is organized. In step 11, the sample set is defined as... ,in yes d 3D eigenvectors Representation and Sample The corresponding set of ambiguous labels, the true label of the sample .in Represents the output space Class tags. Given a candidate label matrix, if the instance Belongs to the tag ( If the membership relationship is non-negative, then the non-negative strength of the membership relationship is... ,otherwise, However, during the training process, Authentic Labels In its ambiguous tag set (Right now In this context, learning algorithms cannot directly access the true values ​​of instances. To address this issue, it's necessary to fully utilize the latent information between instances and candidate labels. Therefore, adjacency graphs are constructed based on the correlation between instance features and candidate labels. The adjacency graph for instance features... If there is an edge between two instances, then ,otherwise The adjacency graph of the label features ,Depend on It is calculated, among which and Examples and The nonnegative membership vectors. In step 12, the purpose of the disambiguation model is to disambiguate the labels of the candidate label set to give them true labels, thus "deceiving" the correction model; the purpose of the correction model is to distinguish correctly disambiguated instances from incorrectly disambiguated instances as much as possible. Specifically, for a given instance, the disambiguation model attempts to fit its potential true label by the distribution of labels on the candidate label set and its connected instances, thus deceiving the correction model, while the correction model attempts to detect whether the disambiguated label is accurate. The specific learning process will be discussed later. Figure 3 The process is described in detail below. In step 13, the learned classifier is used to classify the test samples. Finally, the process proceeds to step 14, the end state. If the results do not meet the pre-defined criteria, the samples need to be filtered, and more samples are selected from the sample storage device for training before returning to step 10. Figure 3 A detailed description of step 12 is provided. Steps 1202-1208 constitute a loop, iteratively cycling between the disambiguation and correction models. These models alternately improve their performance, fully utilizing the potential relationships between instances and candidate labels. Furthermore, the adversarial approach continuously corrects the disambiguated labels, thereby improving the algorithm's accuracy and generalization performance. Specifically, in step 1204, the disambiguation model aims to disambiguate the labels in the candidate label set, giving them true labels to "deceive" the correction model. The correction model aims to distinguish correctly disambiguated instances from incorrectly disambiguated instances as much as possible. Specifically, for a given instance, the disambiguation model attempts to fit its potential true label using the distribution of labels on the candidate label set and its connected instances to deceive the correction model, while the correction model attempts to detect the accuracy of the disambiguated labels, specifically including: (1) Based on equation (1), the function can be maximized and minimized alternately. To learn the optimal parameters for the disambiguation model and the correction model. In each iteration, the disambiguation label obtained through the disambiguation model... Non-candidate labels To train the correction model Among them, the disambiguation model In the correction model Guided by the algorithm, the parameters are updated to approximate the true label distribution from the candidate label set. Iterative competition between the two models drives each other to improve performance until the model output no longer changes.

[0015] Furthermore, in step (4), by correcting instances of incorrect disambiguation detected by the model, the weights of the candidate labels for those instances are adjusted to reduce the impact of incorrect labels, and label disambiguation is performed again. This allows for iterative improvement of model performance through alternating iterations. In addition, when considering the implementation of the disambiguation model, a novel method for calculating the partial label cross-entropy loss is proposed to overcome the limitation that the traditional cross-entropy formula cannot be applied to ambiguous labels. When considering the implementation of label correction, a novel method for setting label probability thresholds is proposed to correct incorrectly disambiguated labels, specifically including: Disambiguation Model: To find the correct label for each training instance, we designed a novel bi-graph convolutional autoencoder partial label learning algorithm. This model has the following capabilities: 1) using the instance adjacency matrix... Adjacency matrix of candidate labels 1) Map instances and candidate label sets to a low-dimensional embedding space; 2) Dynamically update the confidence relationship between instances and candidate labels while learning new representations of instances and label nodes; 3) Achieve ambiguity resolution between instances and candidate label sets by reconstructing the ambiguity of our initialized labels.

[0016] Bi-Graph Convolutional Encoder: Graph convolutional layers contain neighbor information, and the latent information of feature values ​​is passed and transformed between graph edges. We leverage this property by using instance adjacency graphs during propagation. Adjacency graph with labels To obtain more representative embedding representations, we aim to predict link weights by continuously learning node representations from uncertain link relationships. Therefore, dynamically adjusting the propagation weights between nodes and their neighbors by considering the features themselves is essential. To learn rich graph patterns introduced by different edge types, we employ a two-layer Graph Attention Network (GAT) as the message-passing and encoder layers of a dual-graph convolutional encoder, allowing... For the first The output of each convolutional layer Its form is as follows: (2) in It can be an instance or a tag (abbreviated as) or ), and It consists of the instance adjacency matrix and the label adjacency matrix. and . This indicates a multi-head attention mechanism. For learning parameters, The attention coefficient can be calculated using the following formula: (3) To achieve the final embedding of instance nodes and candidate tag nodes, we convert the intermediate output instance encoding and candidate tag encoding as follows: (4) Linear Decoder: For the disambiguation candidate label link weights in the instance graph, we consider using a bilinear decoder, which performs bilinear operations and then applies them. softmax The function generates a probability distribution over possible candidate labels: (5) middle For training parameters, the model employs a non-linear transformation that converts the representations of the instance confidence matrix and the candidate label confidence matrix into a unified embedding representation. After connection, the confidence matrix is ​​passed through a fully connected layer network to obtain the final label confidence matrix, calculated as follows: (6) To optimize the disambiguation model, we minimize the partial label cross-loss function while using a graph Laplacian regularization term in the loss function to smooth out candidate label information. (7) (9) To optimize the disambiguation model, we minimize the partial label cross-loss function while using a graph Laplacian regularization term in the loss function to smooth out candidate label information. (10) middle The proposed method for calculating the partial label cross-entropy loss updates the dissimilarity matrix through iterative calculation. This allows the model to converge optimally to the potential true labels, thereby achieving the purpose of disambiguation. The purpose of this is to smooth the label information so that similar instances in the feature space can share similar labels.

[0017] Positive Model: In contrast to the disambiguation model, the discriminative model aims to maximize the log probabilities of the disambiguation label and non-candidate labels. Traditional graph network computation requires the entire graph as input, which has the following drawbacks: 1) The computational cost increases exponentially with increasing data; 2) The trained model is difficult to apply to newly added data points. Therefore, it is crucial to learn methods to generate effective node embeddings for unknown data using node feature information. This invention overcomes the shortcomings of traditional graph networks by continuously randomly sampling subgraphs from a large graph and approximating the large graph through unbiased estimation. Through graph sampling, a complete discriminative model is constructed to generate embeddings and calculate the loss for each embedding, which can be calculated as follows: (8) middle Yes, the disambiguation label of the subgraph node. Yes, the membership matrix corresponding to the complement of the candidate labels of the subgraph nodes. This corrects the label confidence score of the model output. The probability threshold for the label is calculated using the label confidence score matrix of the corrected model output. (9) middle Then it can be done Correct any potentially misdisambiguated labels, and update the initial confidence function by correcting the model's predictions as follows: (10) middle It is a diagonal matrix. Therefore, the accuracy and generalization performance of the model can be continuously improved by combining the disambiguation labels of the disambiguation model with the label correction of the correction model.

[0018] The disambiguation result of the subsequent instance label is obtained from step 1208.

[0019] Existing machine learning classification techniques all require a large number of accurate labels. This invention can use samples with ambiguous labels for training and predict the test samples, which can greatly reduce the cost of manual labeling, avoid errors in manual labeling, make full use of the potential relationship between instances and candidate labels, and continuously correct the disambiguation labels through the idea of ​​game theory, thereby improving the accuracy and generalization performance of the algorithm.

Claims

1. A weakly supervised classification method based on a graph disambiguation correction network, which specifically includes the following steps: (1) Construct a set of ambiguous labels according to the pre-established rules and select samples from the sample storage device; (2) Extract corresponding features based on the characteristics and features of the samples themselves, and organize the ambiguous label set of the corresponding samples; (3) In this invention, we propose a graph disambiguation correction network. The algorithm consists of two parts: the disambiguation model aims to disambiguate the labels of the candidate label set to make them have true labels, thus "deceiving" the correction model; the correction model aims to distinguish correctly disambiguated instances from incorrectly disambiguated instances as much as possible. Specifically, for a given instance, the disambiguation model attempts to fit its potential true label by the label distribution of the candidate label set and its connected instances, thus deceiving the correction model, while the correction model attempts to detect whether the disambiguated label is accurate. (4) By correcting instances of incorrect disambiguation detected by the model, the weights of the candidate labels for those instances are adjusted to reduce the impact of incorrect labels, and label disambiguation is performed again. In this way, the two processes can be alternately and iteratively improved to enhance the model's performance. Furthermore, when considering the implementation of the disambiguation model, a novel method for calculating the partial label cross-entropy loss is proposed to overcome the limitation that the traditional cross-entropy formula cannot be applied to ambiguous labels. When considering the implementation of label correction, a novel method for setting a label probability threshold is proposed to correct incorrectly disambiguated labels. (5) If the user is satisfied with the prediction result of the previous step, the process ends; otherwise, more samples are selected from the sample storage device for training, and the process proceeds to step (2).

2. The graph disambiguation rectification network based weakly supervised classification method of claim 1, wherein: In step (2), corresponding features are extracted based on the characteristics and features of the samples themselves, and the ambiguous label set of the corresponding samples is organized. Specifically, the sample set is defined as follows: ,in yes d 3D eigenvectors Representation and Sample The corresponding set of ambiguous labels, the true label of the sample. .in Represents the output space Class tags. Given a candidate label matrix, if the instance Belongs to the tag ( If the membership relationship is non-negative, then the non-negative strength of the membership relationship is... ,otherwise, However, during the training process, Authentic Labels In its ambiguous tag set (Right now In this context, learning algorithms cannot directly access the true values ​​of instances. To address this issue, it's necessary to fully utilize the latent information between instances and candidate labels. Therefore, adjacency graphs are constructed based on the correlation between instance features and candidate labels. The adjacency graph for instance features... If there is an edge between two instances, then ,otherwise The adjacency graph of the label features ,Depend on It is calculated, among which and Examples and The non-negative membership vector.

3. The graph disambiguation rectification network based weakly supervised classification method of claim 1, wherein: In step (3), the purpose of the disambiguation model is to disambiguate the labels in the candidate label set, giving them true labels to "deceive" the correction model; the purpose of the correction model is to distinguish correctly disambiguated instances from incorrectly disambiguated instances as much as possible. Specifically, for a given instance, the disambiguation model attempts to fit its potential true label by the distribution of labels on the candidate label set and its connected instances, thus deceiving the correction model, while the correction model attempts to detect whether the disambiguated label is accurate, specifically including: （1） Based on equation (1), the function can be maximized and minimized alternately. To learn the optimal parameters for the disambiguation model and the correction model. In each iteration, the disambiguation label obtained through the disambiguation model... Non-candidate labels To train the correction model Among them, the disambiguation model In the correction model Guided by the algorithm, the parameters are updated to approximate the true label distribution from the candidate label set. Iterative competition between the two models drives each other to improve performance until the model output no longer changes.

4. The weakly supervised classification method based on graph disambiguation correction network according to claim 1, characterized in that: In step (4), the weights of candidate labels for instances that were incorrectly disambiguated by the model are adjusted to reduce the impact of incorrect labels, and label disambiguation is performed again. This iterative process improves model performance. Furthermore, when considering the implementation of the disambiguation model, a novel method for calculating the partial label cross-entropy loss is proposed to overcome the limitation that the traditional cross-entropy formula cannot be applied to ambiguous labels. When considering the implementation of label correction, a novel label probability threshold method is proposed to correct incorrectly disambiguated labels, specifically including: Disambiguation Model: To find the correct label for each training instance, we designed a novel bi-graph convolutional autoencoder partial label learning algorithm. This model has the following capabilities: 1) Using the instance adjacency matrix... Adjacency matrix of candidate labels 1) Map instances and candidate label sets to a low-dimensional embedding space; 2) Dynamically update the confidence relationship between instances and candidate labels while learning new representations of instances and label nodes; 3) Achieve ambiguity resolution between instances and candidate label sets by reconstructing the ambiguity of our initialized labels.

5. Dual-Graph Convolutional Encoder: Graph convolutional layers contain neighbor information, and the latent information of feature values ​​is passed and transformed between graph edges. We leverage this property to use instance adjacency graphs during propagation. Adjacency graph with labels To obtain more representative embedding representations, we aim to predict link weights by continuously learning node representations from uncertain link relationships. Therefore, dynamically adjusting the propagation weights between nodes and their neighbors by considering the features themselves is essential. To learn rich graph patterns introduced by different edge types, we employ a two-layer Graph Attention Network (GAT) as the message-passing and encoder layers of a dual-graph convolutional encoder, allowing... For the first The output of each convolutional layer Its form is as follows: （2） in It can be an instance or a tag (abbreviated as) or ), and It consists of the instance adjacency matrix and the label adjacency matrix. and . This indicates a multi-head attention mechanism. For learning parameters, The attention coefficient can be calculated using the following formula: （3） To achieve the final embedding of instance nodes and candidate tag nodes, we convert the intermediate output instance encoding and candidate tag encoding as follows: （4） Bilinear decoder: For the disambiguation candidate label link weights in the instance graph, we consider using a bilinear decoder that produces a probability distribution over possible candidate labels by bilinearly operating and subsequently applying softmax a softmax function. （5） in For training parameters, the model employs a non-linear transformation that converts the representations of the instance confidence matrix and the candidate label confidence matrix into a unified embedding representation. After connection, the confidence matrix is ​​passed through a fully connected layer network to obtain the final label confidence matrix, calculated as follows: （6） To optimize the disambiguation model, we minimize the partial label cross-loss function while using a graph Laplacian regularization term in the loss function to smooth out candidate label information. （7） in The proposed method for calculating the partial label cross-entropy loss updates the dissimilarity matrix through iterative calculation. This allows the model to converge optimally to the potential true labels, thereby achieving the purpose of disambiguation. The purpose of this is to smooth the label information so that similar instances in the feature space can share similar labels.

6. Correction Model: In contrast to the disambiguation model, the discriminative model aims to maximize the log probabilities of the disambiguated label and the non-candidate label. Traditional graph network computation requires the entire graph as input, which has the following drawbacks: 1) The computational cost increases exponentially with increasing data; 2) The trained model is difficult to apply to newly added data points. Therefore, it is crucial to learn methods to generate effective node embeddings for unknown data using node feature information. This invention overcomes the shortcomings of traditional graph networks by continuously randomly sampling subgraphs from a large graph and approximating the large graph through unbiased estimation. Through graph sampling, a complete discriminative model is constructed to generate embeddings and calculate the loss for each embedding, which can be calculated as follows: （8） in Yes, the disambiguation label of the subgraph node. Yes, the membership matrix corresponding to the complement of the candidate labels of the subgraph nodes. This corrects the label confidence score of the model output. The probability threshold for the label is calculated using the label confidence score matrix of the corrected model output. （9） wherein Then, the labels that can be disambiguated by correction can be identified by The initial confidence function is updated by the correction model's prediction results while correcting the possibly erroneous disambiguated labels as follows: （10） wherein is a diagonal matrix, . Thus, the accuracy and generalization performance of the model can be continuously improved through disambiguation labels of the disambiguation model and label correction of the correction model.

Images