Paper field classification method and device in citation network based on pseudo-label debiasing and medium

Abstract

The application discloses a paper field classification method and device in a citation network based on pseudo-label debiasing and a medium, the method is to obtain the node representation of each view in the citation network for the encoder of the graph neural network; based on the pseudo-label of the semi-supervised node classifier for the data of the unannotated research field category in the citation network, the negative sample is selected, and the contrast learning task is established; the pseudo-label debiasing method is introduced, and the unevenness problem of the paper research field category distribution in the citation network data set is processed; a combined loss function including the classification loss of the label node, the contrast loss of the sampling node and the adaptive boundary loss is constructed, the graph neural network is optimized, the node representation with better discriminability and generalization is learned, and the classification of the paper research field in the citation network is realized. The application can be better applied to a large number of unannotated label citation networks, and the classification of the research field of the paper in the citation network is more accurate and efficient.

Description

Technical Field

[0001] This invention belongs to the field of node classification in complex networks, and specifically relates to a method, apparatus and medium for classifying paper domains in citation networks based on pseudo-label bias removal. Background Technology

[0002] In the problem of classifying the research domain of papers in citation networks, Lauscher et al. were the first to propose applying CNNs to qualitative citation analysis, classifying citations based on both sentiment and purpose. Regarding model structure, they proposed combining word embedding technology with CNN and SVM classification models for performance comparison experiments. In the model design, domain-specific word embeddings were used, meaning that a corpus with the same or similar domain background as the citations to be classified was used during word vector training, and word vectors calculated from a general corpus were cross-referenced to verify the effectiveness of domain-specific word embeddings. However, identifying the research domain of papers in citation networks, which contain a large amount of unlabeled information, is quite difficult. Many classification methods applied to citation networks are often affected by the uneven distribution of categories in the initial dataset, misclassifying papers with scarce labels as belonging to widely distributed categories, thus causing errors in paper classification. Graph contrastive learning is needed to reduce this error.

[0003] In node-level contrastive learning, instance contrastive learning is a classic framework. Its core idea is to distinguish between different augmented samples based on their similarity, making them dissimilar to other samples. This framework has achieved widespread success in computer vision, with typical models such as SimCLR, MoCo, and SimSaim having a profound impact. In the graph domain, these methods have also had a significant influence on contrastive learning. Graph contrastive learning methods all rely on negative samples; however, on large-scale graphs, negative samples lead to huge memory requirements and computational costs. Besides instance contrastive frameworks that augment nodes themselves, many methods define cross-level contrastive schemes based on graph structural information, including node-graph contrastive and node-subgraph contrastive. In this regard, some methods focus on capturing higher-level semantic information through contrastive learning between nodes and the entire graph. On the other hand, node-subgraph contrastive methods focus on introducing subgraph information into node-level contrastives to provide more contextual background. However, due to limitations in memory and computational power, many methods cannot be directly applied to large-scale graph data; and they do not consider the impact of negative sample sampling bias on the model's learned representation. Furthermore, even with methods designed for negative sample selection, the accuracy of negative sample selection cannot be guaranteed. Therefore, constructing accurate negative sample pairs and reducing noise in negative samples is an important research direction at present. Summary of the Invention

[0004] Purpose of the invention: To address the problem of classifying the research domain of papers in citation networks containing semi-supervised nodes in complex networks, this paper proposes a method, device, and medium for classifying the research domain of papers in citation networks based on pseudo-label bias removal. This improves the node representation of the semi-supervised nodes learned by the classification method and enhances the accuracy of node classification.

[0005] Technical solution: The present invention provides a method for classifying paper domains in a citation network based on pseudo-label bias removal, comprising the following steps:

[0006] (1) An encoder based on a graph neural network obtains the node representation of each view in the citation network;

[0007] (2) Negative samples are selected based on pseudo-labels assigned to data in the citation network that are not labeled with the research field category by a semi-supervised node classifier, so as to establish a contrastive learning task;

[0008] (3) Introduce the pseudo-label bias removal method to deal with the uneven distribution of research field categories in the citation network dataset;

[0009] (4) Construct a combined loss function including the classification loss of label nodes, the contrast loss of sampling nodes, and the adaptive boundary loss to optimize the graph neural network so that it learns more discriminative and generalizable node representations, thereby classifying the research fields of papers in the citation network.

[0010] Furthermore, the implementation process of step (1) is as follows:

[0011] By applying graph diffusion convolution to each node in the citation network, we obtain the diffused adjacency matrix S, which defines a continuously weighted graph. Generalized graph diffusion is defined as:

[0012]

[0013] Where S is the adjacency matrix obtained after diffusion, considered as the new comparison view, and T k The transition matrix is ​​obtained through AD. -1 Let A be the adjacency matrix, D be the angle matrix, and θ be the angle matrix. k The weighting coefficients are used; the matrix S is sparsified to obtain a new weighted adjacency matrix.

[0014] The encoder processes the initial features of the various views to obtain new node representations H; the encoder is defined as a function f(·), and the feature extraction process of the encoder is expressed as:

[0015] H = f(X, A)

[0016] Where X is the initial feature matrix of the node; a projection layer is added after the encoder to obtain node Z as the input for the contrastive loss calculation; the projection layer consists of two layers, denoted as g(·):

[0017] Z = g(H) = W 1 σ(W 0 H)

[0018] Where σ is the nonlinear activation operation, W 1 and W 0 These are the parameters that need to be optimized in the two projection layers.

[0019] Furthermore, the implementation process of step (2) is as follows:

[0020] The semi-supervised classifier assigns a pseudo-label to each node. As a criterion for selecting negative samples in contrastive learning; if two nodes have the same pseudo-label, then they do not constitute a negative sample pair; defined The pseudo-labels are defined by `Classifier(·)`, which is the classifier. The expression for calculating the pseudo-labels of a node is shown below:

[0021]

[0022] Where N is the number of nodes and C is the number of node categories. This represents the probability that a node belongs to a certain category;

[0023] In a large-scale citation network graph, a sample of nodes is selected as the object of contrastive learning; the sampling ratio λ is set, that is, P = λ·N nodes are sampled for contrastive learning, where N is the total number of nodes;

[0024] Each training batch requires random sampling of nodes to obtain pseudo-label information for the sampled nodes. Then, comparative learning is conducted using the research field category with the highest probability as the subject of the paper.

[0025] Furthermore, the implementation process of step (3) is as follows:

[0026] A bias-reduction learning approach is used to address the imbalanced distribution of pseudo-labels in the citation network of papers by research domain. Two branches are introduced, each with a different degree of perturbation: a strong perturbation and a weak perturbation. The weak perturbation branch is used to generate pseudo-labels.

[0027] The Debiased Pseudo-Labels method is employed to prevent abrupt changes in bias through approximate bias. This method represents the increase in the probability of a particular class compared to the momentum average, thus serving as a pseudo-label. Its specific representation is as follows:

[0028]

[0029]

[0030] Where α is the weak augmentation function, m∈[0,1] is the momentum coefficient, f(·) represents the logit regression of the node after weak augmentation, and p k It is the probability distribution obtained by the softmax function, where λ represents the debiasing factor, which controls the intensity of the debiasing.

[0031] Furthermore, the implementation process of step (4) is as follows:

[0032] The classification loss employs a semi-supervised training method, using the cross-entropy loss function for labeled training nodes, as shown in the following formula:

[0033]

[0034] Where K represents the number of labeled nodes, C represents the number of categories, and Y is the true value. These are the predicted values; in classification tasks, the model is optimized by minimizing the cross-entropy loss.

[0035] Contrastive Loss: The pairwise objective for each positive sample pair is defined as:

[0036]

[0037] Where P is the number of sampling nodes, j∈S(i) represents the negative sample selected by node i through the pseudo-sample selection strategy, and τ is the temperature parameter; the loss of the other view is defined as l(v i ,u i The overall objective is defined as the average of all positive sample pairs.

[0038]

[0039] A debiased contrastive learning objective is adopted. Assuming the noise ratio in the negative samples is σ, the debiased contrastive loss is expressed as:

[0040]

[0041]

[0042] Use an adaptive edge loss instead of the cross-entropy loss calculated for the sampled samples:

[0043]

[0044] Among them, the adaptive boundary z=f(β(x i )), β indicates strong enhancement;

[0045] Pseudo-labels are generated for unlabeled samples from weakly augmented unlabeled data to supervise the model predictions of the strongly augmented model. Samples with the highest probability below the confidence threshold t are considered unreliable and discarded. The final loss is expressed as:

[0046]

[0047] The model is optimized by jointly training three loss functions; the final model's loss function is:

[0048]

[0049] Where φ>0 and Weighted adjustment parameters used to control the important rows of contrast loss and pseudo-label debiasing loss.

[0050] The device according to the present invention includes a memory and a processor, wherein:

[0051] Memory is used to store computer programs that can run on a processor;

[0052] A processor is configured to, while running the computer program, execute the steps of the paper domain classification method in the citation network based on pseudo-label debiasing as described above.

[0053] The present invention discloses a storage medium storing a computer program, which, when executed by at least one processor, implements the steps of the paper domain classification method in a citation network based on pseudo-label bias removal as described above.

[0054] Beneficial Effects: Compared with existing technologies, the beneficial effects of this invention are as follows: This invention proposes a novel semi-supervised learning framework that combines self-supervised contrastive learning with a semi-supervised setting, fully utilizing the rich information contained in unlabeled nodes; this invention employs multi-view contrastive learning to preserve the semantic information of the graph to the maximum extent, which helps to learn richer representations; by integrating information from multiple views, the model can more comprehensively capture the features of the graph structure, improving the representation learning effect of nodes; this invention samples pseudo-labels of some nodes to alleviate erroneous negative sample selection in contrastive learning, and performs de-biasing learning on pseudo-labels to correct pseudo-label bias, effectively mitigating the impact of pseudo-label bias on model learning. Attached Figure Description

[0055] Figure 1 A flowchart of a paper domain classification method in a citation network based on pseudo-label bias removal;

[0056] Figure 2 To illustrate the difference between the contrastive learning method and the pseudo-label selection method, (a) represents the negative sampling strategy of the node-level contrastive learning method; (b) represents the negative sample sampling strategy based on pseudo-label selection.

[0057] Figure 3 This is the t-SNE embedding graph of the nodes in the Cora dataset. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings.

[0059] like Figure 1 As shown, this invention provides a method for classifying paper domains in citation networks based on pseudo-label bias removal. The specific steps are as follows:

[0060] Step 1: The encoder based on the graph neural network obtains the node representation of each view in the citation network.

[0061] In contrastive learning, since citation networks lack label information, the model learns data representations by studying the similarities or differences between data points. A graph neural network (GNN)-based encoder is used to obtain node representations for each view in the citation network; these representations are then mapped to a specified space through projection layers for calculating the contrastive loss.

[0062] By performing graph diffusion convolution on each node, we obtain the adjacency matrix S of the diffusion matrix, which defines a continuously weighted graph. Generalized graph diffusion is defined as:

[0063]

[0064] Where S is the adjacency matrix obtained after diffusion, considered as the new contrast view, and T is the transition matrix (via AD). -1 Let A represent the adjacency matrix, D represent the angle matrix, and θ be the weighting coefficients. The weighting coefficients are predefined by specific diffusion variables of a particular choice, such as Personalized PageRank (PPR) or the Heatkernel (HK). Their specific representations are shown in the formulas:

[0065]

[0066]

[0067] Where α∈[0,1] is the transmission probability.

[0068] The obtained S is a dense matrix. We can reduce the density of the matrix by setting a threshold to ignore values ​​smaller than the threshold, or by using the topk method to retain the first k edges of each node, thus obtaining a new weighted adjacency matrix.

[0069] After obtaining multiple views, the initial features are processed. Due to the presence of a large amount of noisy data and high-dimensional features, processing with an encoder can effectively reduce the feature dimensionality and extract more effective features for subsequent tasks. Through the encoder, a new node representation H can be obtained. The encoder is defined as a function f(·), and the feature extraction process of the encoder can be represented as:

[0070] H = f(X, A)

[0071] Where X is the feature matrix of the node.

[0072] A projection layer is added after the encoder. After passing through two projection layers, the node representation Z is obtained and used as the input for the contrastive loss calculation. The projection layer is represented as g(·), and the formula for the projection layer is as follows:

[0073] Z = g(H) = W 1 σ(W 0 H)

[0074] The above equation represents a two-layer MLP operation, where σ is a non-linear activation operation. By introducing a projection layer, the differences in node representations can be better captured, which helps improve the effectiveness of contrastive learning.

[0075] Step 2: Select negative samples based on pseudo-labels assigned to data in the citation network that are not labeled with research domain categories by a semi-supervised node classifier to establish a contrastive learning task.

[0076] Figure 2 Figure (a) illustrates the negative sampling strategy of the node-level contrastive learning method, where the central node 1 and all other nodes except itself form negative sample pairs. Figure 2 Table (b) illustrates a negative sample sampling strategy based on pseudo-label selection. The second column in the table represents the maximum probability that a corresponding node belongs to a certain class, and the third column represents the most likely class of that node, i.e., the pseudo-label. It can be seen that the center node 1 and node 6 are very likely to belong to the same class. Therefore, in the pseudo-label selection negative sample strategy, the center node 1 and node 6 do not constitute a negative sample pair. To reduce computational costs, comparative learning is performed on some nodes.

[0077] The semi-supervised classifier assigns a pseudo-label to each node. This serves as the criterion for selecting negative samples in contrastive learning. If two nodes have the same pseudo-label, then they do not constitute a negative sample pair. Definition The pseudo-labels are defined by and Classifier(·) is the classifier. The expression for calculating the pseudo-labels of a node is shown below:

[0078]

[0079] Where N is the number of nodes, and C is the number of node categories. This represents the probability that a node belongs to a certain category.

[0080] In large-scale citation network graphs, sampling all nodes for contrastive learning would place a significant burden on computational resources. To reduce computational and time resource consumption, a subset of nodes is selected for contrastive learning. A sampling ratio λ is set, i.e., P = λ·N nodes are sampled for contrastive learning, where N is the total number of nodes.

[0081] Each training batch involves randomly sampling nodes. After obtaining the pseudo-label information Y_P∈R^(P×C) of the sampled nodes, the nodes with the highest probability are used as the research domain category of the paper for comparative learning.

[0082] Step 3: Introduce pseudo-label bias removal method to address the uneven distribution of research field categories in the citation network dataset.

[0083] Considering that pseudo-labels may be affected by the uneven distribution of research domain categories in the citation network, even if the original data is evenly distributed, the distribution of pseudo-labels learned by the classifier may still be unbalanced. Therefore, a pseudo-label debiasing method is introduced to address the problem of uneven distribution of research domain categories in the citation network dataset.

[0084] A bias-reduction learning approach is used to address the imbalanced distribution of pseudo-labels in the research domains of papers in the citation network. A similar architecture to FixMatch is used, introducing two branches that apply perturbations of different degrees: one strong and one weak. The weak perturbation branch is used to generate pseudo-labels.

[0085] Due to biases in the sample and the model itself, the model often tends to predict most samples as belonging to a few main research domains. To address this bias, the Debiased Pseudo-Labels method is employed. This method prevents abrupt changes in bias by approximating the bias, allowing the model's prediction of a sample to represent the increased probability of a certain class compared to the momentum average, thus serving as a pseudo-label. Its specific representation is as follows:

[0086]

[0087]

[0088] Where α is the weak augmentation function, m∈[0,1] is the momentum coefficient, f(·) represents the logit regression of the node after weak augmentation, and p k It is the probability distribution obtained by the softmax function, where λ represents the debiasing factor, which controls the intensity of the debiasing.

[0089] Step 4: Construct a combined loss function that includes classification loss for label nodes, contrastive loss for sampling nodes, and adaptive boundary loss. Optimize the graph neural network to learn more discriminative and generalizable node representations, enabling the classification of research areas in citation networks. This optimized combination of losses allows for better model training, enabling the model to learn more discriminative and generalizable node representations.

[0090] Classification loss: A semi-supervised training method is used. For labeled training nodes, the cross-entropy loss function is applied, as shown in the following formula:

[0091]

[0092] Where K represents the number of labeled nodes, C represents the number of categories, and Y is the true value. This represents the predicted value. In classification tasks, the model is optimized by minimizing the cross-entropy loss.

[0093] Contrastive Loss: InfoNEC Loss is a common form of loss function in contrastive learning. For sampled nodes, the loss value should be low if the center sample is similar to the positive sample and dissimilar to other negative samples. Conversely, if the center sample is dissimilar to the positive sample or similar to other negative samples, the loss value should be large, thus penalizing the model and prompting parameter updates. The pairwise objective for each positive sample pair is defined as:

[0094]

[0095] Where P is the number of sampling nodes, j∈S(i) represents the negative sample selected by node i through the pseudo-sample selection strategy, and τ is a temperature parameter used to control the degree of attention given to difficult samples. Since the two views are symmetrical, the loss of the other view is defined as l(v i ,u i Finally, the overall objective is defined as the average of all positive sample pairs:

[0096]

[0097] However, the noise from pseudo-labels leads to a bias in the selection of negative samples. To mitigate the impact of negative samples on the model, a bias-reducing contrastive learning objective is adopted. Assuming the noise proportion in the negative samples is σ, the bias-reducing contrastive loss can be expressed as:

[0098]

[0099]

[0100] This invention uses an adaptive edge loss to replace the cross-entropy loss calculated for each sampled sample. In effect, it calculates an adaptive edge loss Δ based on the predicted probabilities of different classes. j The loss is reduced by using adaptive boundaries, and its specific formula is defined as follows:

[0101]

[0102] in z=f(β(x i )), where β represents strong enhancement. In this paper, the diffused view is used as the representation of strong enhancement.

[0103] Furthermore, during the initial training phase, the model is unstable, and unlabeled data can be filtered out. Only data with a predicted probability greater than a predetermined threshold t will be included in the loss calculation. That is, unlabeled data with low probabilities will not contribute to the final loss. Specifically, pseudo-labels for unlabeled samples are generated from weakly augmented unlabeled data to supervise the model predictions of the strongly augmented model. Samples with a maximum probability below the confidence threshold t are considered unreliable and discarded. The final loss is expressed as:

[0104]

[0105] In summary, by jointly training three loss functions, the model is optimized through combination. The final model's loss function can be expressed as:

[0106]

[0107] Where φ>0 and Weighted adjustment parameters used to control the important rows of contrast loss and pseudo-label debiasing loss.

[0108] The present invention also provides an apparatus comprising a memory and a processor, wherein: the memory is used to store a computer program capable of running on the processor; and the processor is used to execute, when running the computer program, the steps of the paper domain classification method in the citation network based on pseudo-label debiasing as described above.

[0109] The present invention also provides a storage medium storing a computer program, which, when executed by at least one processor, implements the steps of the paper domain classification method in a citation network based on pseudo-label bias removal as described above.

[0110] Experimental setup and results, and visualization: This invention was tested on three commonly used citation datasets (i.e., Cora, CiteSeer, PubMed), two co-author datasets (i.e., Coauthor-CS and Coauthor-Physics), and two social network datasets (i.e., Amazon-Photo and Amazon-Computers). Figure 3 As shown, the t-SNE method is used to visualize the embeddings of the Cora dataset generated by different methods. Figure 3 (a) is an embedded graph generated using the method proposed in this patent. Figure 3 In the diagrams (b), (c), (d), (e), and (f), the embedding graphs were generated using the GCN, GAT, MVGRL, GRACE, and BGRL methods, respectively.

[0111] For small citation networks (Cora, PubMed, CiteSeer), training was performed for 200 epochs with a learning rate of 0.005 and weight decay of 5e-4. For large citation networks (Coauthor-CS, Coauthor-Physics), training was performed for 500 epochs with a learning rate of 0.01. For social network datasets, training was performed for 700 epochs with a learning rate of 0.01. The embedding dimension was set to 128 for all datasets. The parameter τ in contrastive learning was set to 0.4, and the loss function included φ and... The parameters were set to 0.8 and 1.3, initially set to 0.001, with the momentum in pseudo-label debiasing set to 0.999, and the proportion of randomly discarded features in weak enhancement selection set to 0.1. Other hyperparameters, such as pseudo-label debiasing probability, threshold, and matrix sparsity, were adjusted for different datasets. For the citation network dataset, a widely used semi-supervised node classification standard was adopted, using 20 nodes per class as the training set, 500 nodes as the validation set, and 1000 nodes as the test set. For the co-author network and social network datasets, the same dataset partitioning strategy as the baseline method was adopted, with 30 nodes per class as the training set, 30 nodes per class as the test set, and the remaining nodes as the test set.

[0112] The model was compared with traditional GCN and GAT methods, as well as the latest comparison methods such as DGI, MMVRL, GRACE, CG3, and SUBG-CON, in node classification on seven real networks. The experimental results are shown in Table 1. OOM indicates insufficient memory on a 16GB GPU, * indicates that the source code cannot be obtained from the paper, and bolded content represents the best accuracy.

[0113] Table 1. Accuracy (%) of node classification

[0114]

[0115] The experiment uses accuracy as the evaluation metric for node classification performance. Accuracy is represented by the ratio of correctly classified samples; a higher value indicates better performance.

[0116] The representations learned by the model were analyzed, and the Cora dataset embeddings generated by different methods were visualized using the t-SNE method. Figure 2 As can be seen, the embedding results projected by the model of this invention show a greater concentration of similar categories and clearer boundaries between categories compared to other methods. Therefore, the proposed model can learn representations that are effective for classification tasks.

[0117] Table 2 Variable Table

[0118]

[0119]

[0120] Table 2 is the variable table involved in this invention. This invention can be combined with a computer system to classify the research fields of papers in a citation network containing semi-supervised nodes.

[0121] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for classifying paper domains in a citation network based on pseudo-label bias removal, characterized in that, Includes the following steps: (1) The encoder based on the graph neural network obtains the node representation of each view in the citation network; (2) Negative samples are selected based on pseudo-labels assigned to data in the citation network that are not labeled with the research field category by a semi-supervised node classifier, so as to establish a contrastive learning task; (3) Introduce the pseudo-label bias removal method to deal with the uneven distribution of research field categories in the citation network dataset; (4) Construct a combined loss function including the classification loss of label nodes, the contrast loss of sampling nodes, and the adaptive boundary loss to optimize the graph neural network so that it learns more discriminative and generalizable node representations, thereby classifying the research fields of papers in the citation network; The implementation process of step (1) is as follows: By performing graph-diffusion convolution on each node in the citation network, a diffused adjacency matrix is ​​obtained. This matrix defines a continuously weighted graph; generalized graph diffusion is defined as: in, The adjacency matrix obtained after diffusion is considered as a new comparison view. The transition matrix is ​​obtained by... definition, It is an adjacency matrix. Represents the angle matrix, These are the weighting coefficients; for the matrix After sparsification, a new weighted adjacency matrix is ​​obtained. ; The encoder processes the initial features of the various views obtained to produce new node representations. Define the encoder as a function. The feature extraction process of the encoder is represented as follows: in, The initial feature matrix of the node is used; a projection layer is added after the encoder to obtain the node. As input for the comparison loss calculation; The implementation process of step (2) is as follows: The semi-supervised classifier assigns a pseudo-label to each node. This serves as the criterion for selecting negative samples in contrastive learning; if two nodes have the same pseudo-label, then they do not constitute a negative sample pair; [Definition] It is a pseudo-tag. As a classifier, the expression for calculating the pseudo-label of a node is shown below: in, For the number of nodes, For the number of node categories, This represents the probability that a node belongs to a certain category; In a large-scale citation network graph, a subset of nodes are selected as the objects of contrastive learning; a sampling ratio is set. That is, sampling Comparative learning is performed on each node. This is the total number of nodes; Each training batch requires random sampling of nodes to obtain pseudo-label information for the sampled nodes. Then, comparative learning is conducted using the research field category with the highest probability as the subject of the paper; Step (4) employs a biased contrastive learning objective, assuming that the noise proportion in the negative samples is... The biased contrast loss is then expressed as: in, The number of nodes sampled. Represents a node Negative samples selected through a spurious sample selection strategy It is a temperature parameter.

2. The paper domain classification method in citation networks based on pseudo-label bias removal according to claim 1, characterized in that, The implementation process of step (3) is as follows: A bias-reduction learning approach is used to address the imbalanced distribution of pseudo-labels in the citation network of papers by research domain. Two branches are introduced, each with a different degree of perturbation: a strong perturbation and a weak perturbation. The weak perturbation branch is used to generate pseudo-labels. The Debiased Pseudo-Labels method is employed to prevent abrupt changes in bias through approximate bias. This method represents the increase in the probability of a particular class compared to the momentum average, thus serving as a pseudo-label. Its specific representation is as follows: in, It is a weakly augmenting function. It is the momentum coefficient. This represents the logit regression after weak augmentation of the nodes, while It is a probability distribution obtained through the softmax function. This represents the debiasing factor, which controls the intensity of the debiasing.

3. The paper domain classification method in citation networks based on pseudo-label bias removal according to claim 1, characterized in that, The implementation process of step (4) is as follows: The classification loss employs a semi-supervised training method, using the cross-entropy loss function for labeled training nodes, as shown in the following formula: in, This represents the number of tagged nodes. Represents the number of categories. For the true value, These are the predicted values; in classification tasks, the model is optimized by minimizing the cross-entropy loss. Contrastive Loss: The pairwise objective for each positive sample pair is defined as: in, The number of nodes sampled. Represents a node Negative samples selected through a spurious sample selection strategy It is a temperature parameter; the loss of another view is defined as... The overall objective is defined as the average of all positive sample pairs: Use an adaptive edge loss instead of the cross-entropy loss calculated for the sampled samples: Among them, the adaptive boundary , , , Indicates strong enhancement; Pseudo-labels are generated from weakly augmented unlabeled data to supervise the model predictions of the strongly augmented model, with the highest probability at the confidence threshold. The following samples are considered unreliable and are discarded. The final loss is expressed as: The model is optimized by jointly training three loss functions; the final model's loss function is: in, and This is a weighted adjustment parameter used to control the important rows of contrast loss and pseudo-label debiasing loss.

4. The paper domain classification method in citation networks based on pseudo-label bias removal according to claim 1, characterized in that, The projection layer consists of two layers, and the projection layer is represented as follows: : in, It is a non-linear activation operation. and These are the parameters that need to be optimized in the two projection layers.

5. A device, characterized in that, Includes memory and processor, wherein: Memory is used to store computer programs that can run on a processor; A processor, configured to, while running the computer program, execute the steps of the paper domain classification method in a citation network based on pseudo-label debiasing as described in any one of claims 1 to 4.

6. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by at least one processor, implements the steps of the paper domain classification method in a citation network based on pseudo-label bias removal as described in any one of claims 1 to 4.

Images