An unsupervised robust graph structure learning method and system based on contrastive learning
By constructing a graph learner and a self-supervised network based on contrastive learning for unsupervised robust graph structure learning, and optimizing the graph structure, we solve the problems of label dependence and poor generalization ability of graph neural networks in unsupervised environments, and achieve stronger graph structure learning and adversarial attack immunity.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF SCI & TECH
- Filing Date
- 2023-06-13
- Publication Date
- 2026-04-14
AI Technical Summary
Existing graph neural network methods for learning graph structures in unsupervised environments are heavily reliant on label information, suffer from edge distribution bias on semi-supervised datasets, exhibit poor generalization ability, and are susceptible to adversarial attacks.
We employ an unsupervised robust graph structure learning method based on contrastive learning. By constructing a graph learner, anchoring graph pairs, data augmentation, and encoder processing, we optimize the graph structure using node feature information and a self-supervised network, and construct a cross-entropy loss function for training to achieve unsupervised learning.
It achieves label-free dependency in an unsupervised environment, reduces model bias, improves the generalization ability of graph structures, and enhances immunity to adversarial attacks.
Smart Images
Figure CN116957048B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machine learning technology, and in particular to an unsupervised robust graph structure learning method and system based on contrastive learning. Background Technology
[0002] Graphs are widely used to represent objects and their complex interactions. As a powerful tool for learning from structured graphical data, graph neural networks have been widely applied to analytical tasks across various fields. The success of graph neural networks can be attributed to their ability to simultaneously utilize the rich information inherent in the structure and properties of graphs. However, inevitably, the graphs provided by graph neural networks are incomplete and noisy, which poses a significant challenge to applying graph neural networks to real-world problems.
[0003] From a representation learning perspective, graph neural networks compute node embeddings by recursively aggregating information from neighboring nodes. This iterative mechanism has a cascading effect—small noises propagate to surrounding nodes, degrading the representation quality of many other nodes. Take social networks as an example, where nodes correspond to users and edges represent friendships. Fraudulent accounts establish fake links to legitimate accounts, thus easily injecting incorrect information into the entire network using the recursive aggregation scheme of graph neural networks, making it difficult to estimate account credibility. Furthermore, recent research shows that imperceptible intentional perturbations in graph structures (i.e., adversarial attacks) can easily lead to incorrect predictions in most graph neural networks. Therefore, high-quality graph structures are typically required for graph neural networks to learn informative representations. Thus, graph structure learning has become a crucial part of graph problems. Most existing graph structure learning methods are conducted in supervised or semi-supervised environments, leading to the following problems: heavy reliance on label information; biased edge distribution on semi-supervised datasets; and poor generalization due to the influence of downstream task supervision information. Summary of the Invention
[0004] The purpose of this invention is to provide an unsupervised robust graph structure learning method based on contrastive learning, which performs graph structure and representation learning in an unsupervised scenario in the face of adversarial attacks.
[0005] The solution to achieve the purpose of this invention is as follows: an unsupervised robust graph structure learning method based on contrastive learning. The graph structure includes the graph's own label information, the node's own feature information, the association information between nodes, and the category label information between nodes. The specific steps are as follows:
[0006] Step 1: Construct a graph learner based on the feature information of the nodes themselves, including a symmetric sparse processing module and a self-supervised network module, and output a graph structure learning view;
[0007] Step 2: Perform low-rank and sparse enhancement on the original graph structure respectively, and output anchor graph pairs;
[0008] Step 3: Compare the anchoring graph obtained in Step 2 with the graph structure learning view obtained in Step 1. Figure 1 Simultaneously perform data augmentation, including edge removal and feature masking, to obtain a processed view;
[0009] Step 4: Put the processed view result obtained in Step 3 into the encoder to obtain the graph node representations corresponding to the anchor map and the processed view, respectively;
[0010] Step 5: Construct a cross-entropy function based on the obtained graph node representation to calculate the contrastive learning loss, and optimize the parameters of the graph learner in Step 1 and the self-supervised network module therein to obtain the final graph structure learner model and graph structure learning view.
[0011] An unsupervised robust graph structure learning system based on contrastive learning includes a graph learner construction module, an anchor graph pair determination module, a data augmentation module, an encoder processing module, and a contrastive learning loss calculation module. In the system:
[0012] The graph learner construction module constructs a graph learner based on the feature information of the nodes themselves, including a symmetric sparse processing module and a self-supervised network module, and outputs a graph structure learning view.
[0013] The anchor graph pair determination module performs low-rank and sparse enhancement on the original graph structure respectively, and outputs anchor graph pairs;
[0014] The data augmentation module combines anchor graph pairs with graph structure learning views. Figure 1 Simultaneously perform data augmentation, including edge removal and feature masking, to obtain a processed view;
[0015] The encoder processing module puts the processed view result into the encoder to obtain graph node representations corresponding to the anchored graph and the processed view, respectively.
[0016] The contrastive learning loss calculation module constructs a cross-entropy function based on the obtained graph node representation to calculate the contrastive learning loss, and optimizes the parameters of the graph learner and its self-supervised network module to obtain the final graph structure learner model and graph structure learning view.
[0017] Compared with the prior art, the significant advantages of this invention are:
[0018] (1) No dependence on label information: In supervised graph structure methods, manually annotated labels are very scarce and very time-consuming. This independence from the dependence on labels makes this network applicable to more general cases.
[0019] (2) Model bias in unsupervised environment: Node classification usually adopts a semi-supervised environment, that is, only a special part of the nodes have supervised labels. This imbalance leads to the bias of edge distribution, which affects the quality of the learning structure. This model is built in an unsupervised environment, so it can obtain a better learning structure.
[0020] (3) No restrictions on downstream tasks: In existing methods, the structure is learned specifically for node classification, so it may contain more information related to specific tasks rather than general knowledge. Therefore, the generalization ability of the optimized learning structure is poor. This method does not use any supervision information for specific tasks, and the learned graph structure has a stronger generalization ability. Attached Figure Description
[0021] Figure 1 This is a flowchart of the unsupervised robust graph structure learning method based on contrastive learning, which is the subject of this invention. Detailed Implementation
[0022] This invention provides an unsupervised robust graph structure learning method based on contrastive learning. The graph structure includes its own label information, node's own feature information, association information between nodes, and category label information between nodes. The method optimizes graph structures that have been attacked and corrupted. The target of the optimization is the corrupted graph structure. Through this method, an optimized clean graph structure is obtained, which is immune to adversarial attacks. The specific steps are as follows:
[0023] Step 1: Construct a graph learner based on the feature information of the nodes themselves, including a symmetric sparse processing module and a self-supervised network module, and output a graph structure learning view. The self-supervised network module is an original structure update method that will provide more supervision information for the graph learner.
[0024] Step 2: Perform low-rank and sparse augmentation on the original graph structure respectively, and output anchored graph pairs. Low-rank and sparse augmentation is an original graph contrastive learning data augmentation method.
[0025] Step 3: Compare the anchoring graph obtained in Step 2 with the graph structure learning view obtained in Step 1. Figure 1 Simultaneously perform data augmentation, including edge removal and feature masking, to obtain a processed view;
[0026] Step 4: Put the processed view result obtained in Step 3 into the encoder to obtain the graph node representations corresponding to the anchor map and the processed view, respectively;
[0027] Step 5: Construct a cross-entropy function based on the obtained graph node representation to calculate the contrastive learning loss, and optimize the parameters of the graph learner in Step 1 and the self-supervised network module therein to obtain the final graph structure learner model and graph structure learning view.
[0028] As a specific example, the graph structure that needs to be processed is a social network or a literature citation network.
[0029] As a specific example, step 1 involves constructing a graph learner based on the node's own feature information, including a symmetric sparsity processing module and a self-supervised network module, outputting a graph structure learning view, as detailed below:
[0030] FGP learner: Models each element of the adjacency matrix directly using independent parameters, requiring no additional input. An FGP learner is defined as follows:
[0031]
[0032] Where θ = Ω∈R n×n It is a parameter matrix; (·) is a non-linear function that makes training more stable; The learned graph structure matrix; It is a learnable parameter matrix; R n×n It is the set of all n×n dimensional real matrices, where n represents the matrix dimension; the FGP learner assumes that each edge in the graph is independent;
[0033] The self-supervised task is based on the following assumptions: a graph structure suitable for predicting node features is also suitable for predicting node labels; by masking some input features or adding noise to the input features, an independent GNN is trained, and after updating the adjacency matrix, the adjacency matrix can recover the masked features or noise, so that the graph learner can learn a better graph structure; the self-supervised network module is based on denoising autoencoders, where the GNNS receives the generated... The feature matrix with added noise is used to output a new feature matrix, which is then combined with the original feature matrix to obtain the loss function. as follows:
[0034]
[0035] in, The feature matrix after adding noise; This is the original characteristic matrix; It is an adjacency matrix; The feature matrix predicted by GNNS after adding noise and the adjacency matrix; For GNNS parameters; This represents the function for calculating the difference between two feature matrices.
[0036] As a specific example, step 2 involves performing low-rank and sparse enhancements on the original graph structure to output anchored graph pairs, as follows:
[0037] (2.1) Perform SVD decomposition on the graph structure adjacency matrix of the set perspective into the form of multiplication of eigenvector matrix and eigenvalue matrix, delete a set number of singular values greater than a set threshold, and then combine the eigenvector matrix and eigenvalue matrix to obtain the low-rank processed graph structure adjacency matrix S1.
[0038] (2.2) Reduce the nuclear norm of the adjacency matrix so that the contrastive learner can capture the sparsity properties and learn a graph-structured adjacency matrix S2 that is immune to the effects of attacks and has a sparse structure.
[0039] (2.3) The learner's perspective is compared with two different anchor graphs, S1 and S2, to obtain two different contrastive losses. The two losses are then propagated forward to update all parameters of the graph structure learner model. Thus, the entire graph structure learner will be trained from these two directions. After training, two different graph structures will be learned, and these two graph structures will be processed as follows:
[0040]
[0041] in To learn the final graph structure, For anchoring view Figure 1 , For anchoring view 2, To adjust the anchor view weight parameters.
[0042] As a specific example, step 3 involves learning the view by comparing the anchored graph obtained in step 2 with the graph structure obtained in step 1. Figure 1 Simultaneously perform data augmentation, including edge removal and feature masking, to obtain the processed view, as follows:
[0043] (3.1) In order to scramble the node features, a portion of the feature dimensions are randomly selected and masked with zeros; for a given feature matrix X, a masking vector m is first sampled. (x) ∈{0,1} d Where x is a one-dimensional eigenvector, d is the dimension of the eigenvector, and each element is independently derived from a Bernoulli distribution with probability p. (x) Sampling is performed; then, the model uses m (x) Mask the feature vector of each node;
[0044] (3.2) The edge removal module corrupts the graph structure by randomly deleting a portion of the edges; specifically, for a given adjacency matrix A, a masking matrix M is first sampled. (a) ∈{0, 1} n×n Each element Independently from the Bernoulli distribution with probability p (a)Sampling is performed, where a is a one-dimensional feature vector and n×n is the dimension of the adjacency matrix; then, the adjacency matrix is processed using M... (a) Cover it up;
[0045] (3.3) Combine these two augmentation schemes with existing low-rank and sparse augmentation schemes to generate augmentation graph structures on the learner view and anchor view:
[0046]
[0047]
[0048] in l and a These are the enhanced learner view and the anchored view, respectively. and These are the adjacency matrices of the learner view without enhancement and the anchor view without enhancement, respectively. For edge removal transformation, For feature perturbation transformation, For low-rank transformation, It is a sparse transformation;
[0049] To obtain different contexts in the two views, feature masking in the two views uses different probabilities p. l ≠ p a For edge deletion, since the adjacency matrices of the two views are significantly different, the same deletion probability p is used. l =p a = p; For low-rank augmentation and sparse augmentation, a view will be affected simultaneously with a set probability, so that the graph structure learner model is affected by both clean and polluted views at the same time.
[0050] As a specific example, step 4 involves feeding the processed view result obtained in step 3 into the encoder to obtain graph node representations corresponding to the anchor map and the processed view, respectively, as follows:
[0051] (4.1) Encoder based on graph structure learner model (·) From the augmented learner view l and anchored view a Extracting node-level representations:
[0052]
[0053] Where θ is the encoder f θ The parameters of (·); and ∈R n×d1 d1 represents the node representation matrices of the learner view and the anchor view, respectively; d2 represents the dimension. Encoded representation of the learner's view. R is the encoded representation of the anchored view. n ×d1 Given the set of all matrices of dimension n×d1, where n is the number of nodes; the graph structure learner model uses a GCN encoder with its layer number L1 set to 2.
[0054] (4.2) After the encoder, a projector with L2 MLP layers (·) maps the representation to another latent space, where the contrastive loss is computed:
[0055]
[0056] in It is a projector The parameter of (·), and Z l and Z a ∈R n×d2 It is the projection node representation matrix of the learning view / anchored view; where d2 is the projection dimension. For the projected representation of the learner's view, R is a projected representation of the learner's view. n×d2 It is the set of all matrices of dimension n×d2, where n is the number of nodes.
[0057] As a specific example, step 5 involves constructing a cross-entropy function based on the obtained graph node representation to calculate the contrastive learning loss, and optimizing the parameters of the graph learner from step 1 and its self-supervised network module to obtain the final graph structure learner model and graph structure learning view. The specific formula is as follows:
[0058]
[0059]
[0060] in, It is the total loss. Cosine similarity function It's a temperature parameter. Let cross-entropy be the loss function. Let i be the representation of the i-th point in the learner's view. Let k be the representation of the i-th point in the anchored view, and k be the loop parameter, traversing all points in the graph. The graph learner parameters and self-supervised network parameters are obtained by optimizing this function, and finally the learning view is obtained as the optimized robust graph structure.
[0061] This invention also provides an unsupervised robust graph structure learning system based on contrastive learning, including a graph learner construction module, an anchor graph pair determination module, a data augmentation module, an encoder processing module, and a contrastive learning loss calculation module, used to implement the aforementioned unsupervised robust graph structure learning method based on contrastive learning. In the system:
[0062] The graph learner construction module constructs a graph learner based on the feature information of the nodes themselves, including a symmetric sparse processing module and a self-supervised network module, and outputs a graph structure learning view.
[0063] The anchor graph pair determination module performs low-rank and sparse enhancement on the original graph structure respectively, and outputs anchor graph pairs;
[0064] The data augmentation module combines anchor graph pairs with graph structure learning views. Figure 1 Simultaneously perform data augmentation, including edge removal and feature masking, to obtain a processed view;
[0065] The encoder processing module puts the processed view result into the encoder to obtain graph node representations corresponding to the anchored graph and the processed view, respectively.
[0066] The contrastive learning loss calculation module constructs a cross-entropy function based on the obtained graph node representation to calculate the contrastive learning loss, and optimizes the parameters of the graph learner and its self-supervised network module to obtain the final graph structure learner model and graph structure learning view.
[0067] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0068] Example 1
[0069] Combination Figure 1 This invention discloses an unsupervised robust graph structure learning method based on contrastive learning. The graph structure includes the graph's own label information, the node's own feature information, the association information between nodes, and the category label information between nodes. The representation process is as follows:
[0070] Step 1: Construct a graph learner based on the feature information of the nodes themselves, including a symmetric sparse processing module and a self-supervised update module, and output a graph structure learning view;
[0071] Step 2: Perform low-rank and sparse enhancement on the original graph structure respectively, and output anchor graph pairs;
[0072] Step 3: Anchoring the image to the learning view Figure 1 Simultaneously perform data augmentation, including edge removal and feature masking;
[0073] Step 4: Input the processed view into the encoder to obtain the corresponding graph node representation;
[0074] Step 5: Construct a cross-entropy function based on the obtained graph node representation to calculate the contrastive learning loss, and optimize the parameters of the graph learner and the self-supervised network to obtain the final model and graph structure learning view.
[0075] Furthermore, the network is a social network or a document citation network.
[0076] Furthermore, step 1 involves constructing a graph learner based on the node's own feature information, including a symmetric sparsity processing module and a self-supervised update module, outputting a graph structure learning view, as detailed below:
[0077] FGP learners model each element of the adjacency matrix directly with independent parameters, without requiring additional input. Specifically, an FGP learner is defined as:
[0078]
[0079] Where θ = Ω∈R n×n It is a parameter matrix. (·) is a non-linear function, which makes training more stable. The FGP learner assumes that each edge in the graph exists independently.
[0080] The self-supervised task is based on the assumption that a graph structure suitable for predicting node features is also suitable for predicting node labels. It works by masking some input features (or adding noise to them) and training a separate GNN to update the adjacency matrix so that it can recover the masked (or noisy) features. This module is based on Denoising Autoencoders, where the GNN... S The system takes the generated S and the feature matrix with added noise, outputs a new feature matrix, and compares it with the original feature matrix to obtain the loss function as follows:
[0081]
[0082] in This is the feature matrix after adding noise.
[0083] Furthermore, the low-rank and sparse enhancement described in step 2 outputs anchor graph pairs, as follows:
[0084] (1) Perform SVD decomposition on the graph structure adjacency matrix from a certain perspective, delete a certain number of large singular values, and then combine them to obtain the low-rank processed graph structure adjacency matrix S1.
[0085] (2) Reduce the nuclear norm of the adjacency matrix so that the contrastive learner can capture the sparsity properties and learn a graph structure adjacency matrix S2 that is immune to the effects of attacks to a certain extent.
[0086] (3) The learning perspective is compared with two different anchor maps to obtain two different contrastive losses. These two losses can be forward-propagated to update all network parameters. Thus, the entire network will be trained from these two directions. After training, two different graph structures will be learned, and these two graph structures will be processed as follows:
[0087]
[0088] Furthermore, the anchoring map described in step 3 is related to the learning view. Figure 1 Simultaneously, data augmentation is performed, including edge removal and feature masking, as detailed below:
[0089] (1) In order to disrupt the node features, a portion of the feature dimensions are randomly selected and masked with zeros. Specifically, for a given feature matrix X, a masking vector m is first sampled. (x) ∈{0,1} d Each element is independently derived from a Bernoulli distribution with probability p. (x) Sampling is performed. Then, the model uses m (x) Mask the feature vector of each node.
[0090] (2) The model corrupts the graph structure by randomly deleting a portion of the edges. Specifically, for a given adjacency matrix A, a masking matrix M is first sampled. (a) ∈{0, 1} n×n Each element Independently from the Bernoulli distribution with probability p (a) Sampling is performed. Then, the adjacency matrix is processed using M. (a) Cover it up.
[0091] (3) Combine these two augmentation schemes with existing low-rank and sparse augmentation schemes to generate augmentation graph structures on the learner view and anchor view:
[0092]
[0093]
[0094] in l and a These are the enhanced learner view and the anchored view, respectively. To obtain different contexts in the two views, the feature masking for the two views uses different probabilities p. l ≠ p a For edge deletion, since the adjacency matrices of the two views are significantly different, the model uses the same deletion probability p. l= p a = p. For low-rank augmentation and sparse augmentation, the model will simultaneously influence a certain view with a certain probability, so that the model is simultaneously affected by both clean and polluted views.
[0095] Further, step 4 involves inputting the processed view into the encoder to obtain the corresponding graph node representation, as detailed below:
[0096] (1) Encoder f based on graph neural network θ (·) From the enhanced graph l and a Extracting node-level representations:
[0097]
[0098] Where θ is the encoder f θ The parameter of (·), and H l and H a ∈R n×d1 (where d1 represents the dimension) are the node representation matrices for the learner view and the anchor view, respectively. The model uses a GCN encoder with its number of layers L1 set to 2.
[0099] (2) The location is after the encoder, a projector with L2 MLP layers. (·) maps the representation to another latent space, where the contrastive loss is computed:
[0100]
[0101] in It is a projector The parameter of (·), and Z l and Z a ∈Rn×d2 (where d2 is the projection dimension) is the projection node representation matrix of the learned view / anchored view.
[0102] Further, step 5 constructs a cross-entropy function based on the obtained graph node representations to calculate the contrastive learning loss, and optimizes the parameters of the graph learner and the self-supervised network to obtain the final model and graph structure learning view, as follows:
[0103]
[0104]
[0105] Where L is the total loss, sim(·,·) is the cosine similarity function, and t is the temperature parameter. The graph learner parameters and self-supervised network parameters are obtained through optimization of this function, ultimately yielding the learning perspective as the optimized robust graph structure.
[0106] Example 2
[0107] This embodiment uses a text classification task to further illustrate the invention:
[0108] (1) This embodiment uses a standard citation network dataset— Use the dataset to test the dataset.
[0109] The dataset contains 2708 nodes, each representing a paper categorized into 7 classes. Each node has 1433 features, each corresponding to a single word, with a value of either 0 or 1. A value of 0 indicates the word is not mentioned in a paper, while a value of 1 indicates it is. If two papers are cited from each other, there is an edge between them. The data is presented using an adjacency matrix. and characteristic matrix Label vector Formal representation.
[0110] (1) According to the formula The generative learning perspective, the self-supervised loss is obtained by the following formula:
[0111]
[0112] in This is the feature matrix after adding noise.
[0113] (2) Process the obtained anchoring plots using the following formula:
[0114]
[0115] in The balancing coefficient is used to balance the relationship between anchored view pairs. The anchored view and the learning view are enhanced as follows to obtain... l and a :
[0116]
[0117]
[0118] (3) The enhanced view l and aThe following encoding and projection are performed to obtain Z. l and Z a :
[0119]
[0120]
[0121] (4) Z l and Z a The contrast loss is calculated using the following formula:
[0122]
[0123]
[0124] Where L is the total loss, t is the temperature parameter, and is set to 0.2. The Adam algorithm is used to optimize the total loss, and the learning rate is set to 0.001 to obtain the final learning viewpoint graph structure representation.
[0125] In summary, this invention utilizes the feature information between nodes to construct the learner view; it strengthens the graph learner using a self-supervised model, resulting in a more expressive graph structure; it performs low-rank and sparse enhancements on the original graph structure to form anchor graph pairs, which are then fed into the encoder and projector along with the learning viewpoint; it constructs an objective function based on contrastive loss; and it optimizes network parameters through end-to-end joint learning to obtain an optimized graph structure representation. Training in an unsupervised environment broadens its applicability, and the resulting graph structure exhibits stronger generalization ability for downstream tasks.
Claims
1. An unsupervised robust graph structure learning method based on contrastive learning, characterized in that, The graph structure to be processed is a citation network, used to complete a text classification task. The graph structure includes the graph's own label information, the node's own feature information, the association information between nodes, and the category label information between nodes. The specific steps are as follows: Step 1: Construct a graph learner based on the feature information of the nodes themselves, including a symmetric sparse processing module and a self-supervised network module, and output a graph structure learning view; Step 2: Perform low-rank and sparse enhancement on the original graph structure respectively, and output anchor graph pairs; Step 3: Perform data augmentation on the anchored graph obtained in Step 2 and the graph structure learning view obtained in Step 1, including edge removal and feature masking, to obtain the processed view. Step 4: Put the processed view result obtained in Step 3 into the encoder to obtain the graph node representations corresponding to the anchor map and the processed view, respectively; Step 5: Construct a cross-entropy function based on the obtained graph node representation to calculate the contrastive learning loss, and optimize the parameters of the graph learner in Step 1 and the self-supervised network module therein to obtain the final graph structure learner model and graph structure learning view. Step 1 describes constructing a graph learner based on the node's own feature information, including a symmetric sparsity processing module and a self-supervised network module, and outputting a graph structure learning view, as detailed below: FGP learner: Models each element of the adjacency matrix directly using independent parameters, requiring no additional input. An FGP learner is defined as follows: ； Where θ = Ω∈R n×n It is a parameter matrix; (·) is a non-linear function that makes training more stable; The learned graph structure matrix; It is a learnable parameter matrix; R n×n It is the set of all n×n dimensional real matrices, where n represents the matrix dimension; the FGP learner assumes that each edge in the graph is independent; The self-supervised task is based on the following assumptions: a graph structure suitable for predicting node features is also suitable for predicting node labels; by masking some input features or adding noise to the input features, an independent GNN is trained, and after updating the adjacency matrix, the adjacency matrix can recover the masked features or noise; the self-supervised network module is based on a denoising autoencoder, where the GNNS receives the generated... The feature matrix with added noise is used to output a new feature matrix, which is then combined with the original feature matrix to obtain the loss function. as follows: ； in, The feature matrix after adding noise; This is the original characteristic matrix; It is an adjacency matrix; The feature matrix predicted by GNNS after adding noise and the adjacency matrix; For GNNS parameters; This represents the function for calculating the difference between two feature matrices; Step 2 involves performing low-rank and sparse enhancements on the original graph structure to output anchored graph pairs, as detailed below: (2.1) Perform SVD decomposition on the graph structure adjacency matrix of the set perspective into the form of multiplication of eigenvector matrix and eigenvalue matrix, delete a set number of singular values greater than a set threshold, and then combine the eigenvector matrix and eigenvalue matrix to obtain the low-rank processed graph structure adjacency matrix S1. (2.2) Reduce the nuclear norm of the adjacency matrix so that the contrastive learner can capture the sparsity properties and learn a graph-structured adjacency matrix S2 that is immune to the effects of attacks and has a sparse structure. (2.3) The learner's perspective is compared with two different anchor graphs, S1 and S2, to obtain two different contrastive losses. The two losses are then propagated forward to update all parameters of the graph structure learner model. Thus, the entire graph structure learner will be trained from these two directions. After training, two different graph structures will be learned, and these two graph structures will be processed as follows: ； in To learn the final graph structure, For anchoring view 1, For anchoring view 2, To adjust the anchor view weight parameters; Step 3 involves performing data augmentation on the anchored graph obtained in Step 2 and the graph structure learning view obtained in Step 1, including edge removal and feature masking, to obtain the processed view, as detailed below: (3.1) In order to scramble the node features, a portion of the feature dimensions are randomly selected and masked with zeros; for a given feature matrix X, a masking vector m is first sampled. (x) ∈{0,1} d Where x is a one-dimensional eigenvector, d is the dimension of the eigenvector, and each element is independently derived from a Bernoulli distribution with probability p. (x) Sampling is performed; then, the model uses m (x) Mask the feature vector of each node; (3.2) The edge removal module corrupts the graph structure by randomly deleting a portion of the edges; specifically, for a given adjacency matrix A, a masking matrix M is first sampled. (a) ∈{0, 1} n×n Each element Independently from the Bernoulli distribution with probability p (a) Sampling is performed, where a is a one-dimensional feature vector and n×n is the dimension of the adjacency matrix; then, the adjacency matrix is processed using M... (a) Cover it up; (3.3) Combine these two augmentation schemes with existing low-rank and sparse augmentation schemes to generate augmentation graph structures on the learner view and anchor view: ； ； in l and a These are the enhanced learner view and the anchored view, respectively. and These are the adjacency matrices of the learner view without enhancement and the anchor view without enhancement, respectively. For edge removal transformation, For feature perturbation transformation, For low-rank transformation, It is a sparse transformation; To obtain different contexts in the two views, feature masking for the two views uses different probabilities; for edge removal, since the adjacency matrices of the two views are already significantly different, the same removal probability is used; for low-rank enhancement and sparse enhancement, one view is simultaneously affected with a set probability, so that the graph structure learner model is affected by both clean and contaminated views at the same time. Step 4 involves feeding the processed view result obtained in Step 3 into the encoder to obtain graph node representations corresponding to the anchor map and the processed view, respectively, as follows: (4.1) Encoder based on graph structure learner model (·) From the augmented learner view l and anchored view a Extracting node-level representations: ； Where θ is the encoder f θ The parameters of (·); and ∈R n×d1 d1 represents the node representation matrices of the learner view and the anchor view, respectively; d2 represents the dimension. Encoded representation of the learner's view. R is the encoded representation of the anchored view. n×d1 Given the set of all matrices of dimension n×d1, where n is the number of nodes; the graph structure learner model uses a GCN encoder with its layer number L1 set to 2. (4.2) After the encoder, a projector with L2 MLP layers (·) maps the representation to another latent space, where the contrastive loss is computed: ； in It is a projector The parameter of (·), and Z l and Z a ∈R n×d2 It is the projection node representation matrix of the learning view / anchored view; where d2 is the projection dimension. For the projected representation of the learner's view, R is a projected representation of the learner's view. n×d2 Let be the set of all matrices of dimension n×d2, where n is the number of nodes; Step 5 involves constructing a cross-entropy function based on the obtained graph node representations to calculate the contrastive learning loss, and optimizing the parameters of the graph learner from Step 1 and its self-supervised network module to obtain the final graph structure learner model and graph structure learning view. The specific formulas are as follows: ； ； in, It is the total loss. Cosine similarity function It's a temperature parameter. Let cross-entropy be the loss function. Let i be the representation of the i-th point in the learner's view. Let k be the representation of the i-th point in the anchored view, and k be the loop parameter, traversing all points in the graph. The graph learner parameters and self-supervised network parameters are obtained by optimizing this function, and finally the learning view is obtained as the optimized robust graph structure.
2. An unsupervised robust graph structure learning system based on contrastive learning, characterized in that, The system includes a graph learner construction module, an anchor graph pair determination module, a data augmentation module, an encoder processing module, and a contrastive learning loss calculation module, used to implement the unsupervised robust graph structure learning method based on contrastive learning as described in claim 1. In the system: The graph learner construction module constructs a graph learner based on the feature information of the nodes themselves, including a symmetric sparse processing module and a self-supervised network module, and outputs a graph structure learning view. The anchor graph pair determination module performs low-rank and sparse enhancement on the original graph structure respectively, and outputs anchor graph pairs; The data augmentation module performs data augmentation on the anchor graph pairs and the graph structure learning view together, including edge removal and feature masking, to obtain the processed view; The encoder processing module puts the processed view result into the encoder to obtain graph node representations corresponding to the anchored graph and the processed view, respectively. The contrastive learning loss calculation module constructs a cross-entropy function based on the obtained graph node representation to calculate the contrastive learning loss, and optimizes the parameters of the graph learner and its self-supervised network module to obtain the final graph structure learner model and graph structure learning view.