Classification method and system for same and different matching network nodes for graph contrast learning based on macro and micro message passing, and medium

Abstract

The invention discloses a same and different matching network node classification method and system for graph comparison learning based on macro and micro message passing, and a medium, and the method comprises the steps: firstly constructing a k neighbor graph based on node feature similarity through the node features of an original graph, and carrying out the random connection edge removal of the original graph and the k neighbor graph, generating a corresponding structure view and a feature view, then performing view role definition on the structure view and the feature view according to the same and different matching types of the network, and determining a priority view and an auxiliary view; performing optimal and auxiliary view alignment coding on the priority view and the auxiliary view to obtain node representations under different views; further constructing an InfoNCE comparison loss function, and performing optimization training on node embedding representation of the model; and finally, the optimized node embedding representation is used for a node classification task, and training and evaluation are carried out. According to the method, the node representation quality can be effectively improved, so that the accuracy of node classification in the same and different matching network is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of graph contrast learning, and in particular to a method, system and medium for classifying nodes in a same-different matching network for graph contrast learning based on macro-micro message transmission. Background Art

[0002] As a self-supervised representation learning method for graph data, graph contrastive learning has made significant progress in recent years in reducing the reliance on expensive and time-consuming labeled data. The core idea of ​​this technology is to learn effective node or graph representations by constructing a contrastive loss function by maximizing the similarity of positive sample pairs and the difference of negative sample pairs. Generally, graph contrastive learning includes three main steps: view enhancement, view encoding, and contrastive training. Existing studies have significantly improved the performance of graph contrastive learning in downstream graph tasks such as node classification by developing new data enhancement or model enhancement strategies and optimizing contrast objectives.

[0003] Although existing graph contrastive learning methods perform well on homogeneous networks, they generally rely on graph neural network (GNN) encoders in the view encoding process, and these encoders are usually based on the homogeneity assumption, which makes these methods less effective when dealing with heterogeneous networks. The so-called homogeneity assumption means that nodes with the same label tend to be closely connected, but this assumption is often not true in heterogeneous networks, making it difficult for existing methods to adapt to the characteristics of heterogeneous networks. For example, social networks are usually homogeneous, and individuals tend to establish connections based on the same interests or backgrounds; while in protein structure graphs, different types of amino acids tend to be connected together, showing heterogeneity. This assumption limits the direction (macro level) and method (micro level) of message transmission to a certain extent, weakening the performance of graph contrastive learning in dealing with heterogeneous graphs.

[0004] Specifically, at the macro level, homogeneous networks mainly rely on structural neighbors for message passing, while heterogeneous networks rely more on feature neighbors based on node feature similarity. However, most existing methods fail to fully recognize this difference and still use structural neighbors to deal with heterogeneous networks, resulting in poor results. At the micro level, traditional GNNs use nodes themselves as neighbors for feature propagation, but this approach of treating nodes themselves and neighbors equally is difficult to adapt to the characteristics of heterogeneous networks, because in heterogeneous networks, the feature differences between nodes and neighbors are often large. Although some studies have tried to solve this problem by diversifying high-order neighbor aggregation strategies, these methods usually increase algorithmic complexity and may weaken the ability to process homogeneous networks.

[0005] In summary, although there are existing works dedicated to improving the performance of graph contrastive learning in homogeneous and heterogeneous networks, due to insufficient understanding of the differences in the macro-direction and micro-methods of message transmission between these two types of networks, existing methods are still limited in terms of computational cost and practical effect. Therefore, developing a more efficient graph contrastive learning method that encourages the model to fully learn the important features of homogeneous and heterogeneous networks and improves the performance of graph contrastive learning in downstream graph tasks such as node classification has become one of the urgent problems to be solved in the current field. Summary of the invention

[0006] The purpose of the present invention is to address the impact of the homogeneity assumption on the classification task of homogeneous and heterogeneous network nodes, and to propose a homogeneous and heterogeneous network node classification method, system and medium based on macro-micro message passing for graph comparative learning.

[0007] The objective of the present invention is achieved through the following technical solutions:

[0008] In a first aspect, a method for classifying nodes in a homogeneous network based on macro-micro message passing for graph contrast learning is provided, comprising the following steps:

[0009] S1: Based on the original image Where A is the adjacency matrix, X is the node feature matrix, the cosine distance between node features is calculated, and a k-nearest neighbor graph based on the similarity of node features is constructed;

[0010] S2: Perform a view enhancement operation of randomly removing edges on the original graph and the k-nearest neighbor graph to generate the corresponding structural view v str and feature view v att ;

[0011] S3: According to the edge homogeneity ratio h of the original graph edge Determine the type of network configuration and then define the priority view v of the network configuration p Is the structure view v str , auxiliary view v s is the feature view v att ; Priority view v of heterogeneous networks p is the feature view v att , auxiliary view v s Is the structure view v str ;

[0012] S4: Use two independent ASP-SGC encoders to encode the priority view v p and auxiliary view v s Perform priority view alignment coding to obtain the node representation H after priority view coding p The node representation H after alignment encoding with the auxiliary view sa ;

[0013] S5: Node-based representation H p and H sa , define the anchor node v i The InfoNCE contrast loss function is constructed based on the positive and negative sample pairs, and the node embedding representation of the model is optimized and trained through this loss function.

[0014] S6: The optimized node embedding representation is input into the logistic regression node classifier for training and evaluation.

[0015] Preferably, in S1, the cosine distance calculation formula between node features is:

[0016]

[0017] where x i and x j The nodes v i and v j According to the cosine distance calculation result, k nearest neighbor nodes are selected for each node and corresponding edges are added to build a k-nearest neighbor graph based on node feature similarity.

[0018] As a preference, in S3, the edge homogeneity ratio h, which measures the homogeneity level of the network, is edge The calculation formula is:

[0019]

[0020] like but For heterogeneous networks, if but The same network.

[0021] As a preference, in S4, the ASP-SGC encoder uses an adaptive self-propagation strategy to decompose node self-loop feature propagation into neighbor assimilation propagation and self-independent propagation:

[0022]

[0023] The first term (1-ω)H (k) represents the self-independent propagation part, which emphasizes the independence of self-loop feature propagation and does not depend on the neighbor node features; the second term represents the neighbor assimilation propagation part, in which the self-loop acts as a neighbor to propagate features; k represents the node of the kth layer, representing H (k) , ω is the weight coefficient for self-independent propagation, and Θ is the trainable weight matrix.

[0024] As a preference, in S4, the priority view v is first used p and auxiliary view v sEach ASP-SGC encoder encodes the corresponding view and obtains the node representation H p and H s ; Then the ASP-SGC encoder of the auxiliary view is used to encode the priority view v from a global perspective p and compare it with the initial node representation v of the auxiliary view s Add together to generate the aligned auxiliary view and the final node representation H sa :

[0025]

[0026] As a preferred embodiment, in S5, an anchor node v is defined i The positive sample pair is Negative sample pairs include as well as The InfoNCE contrast loss is constructed as:

[0027]

[0028] where τ is the temperature parameter and s(·) is the cosine similarity.

[0029] In a second aspect, a system for classifying nodes in a same-different network using graph contrast learning based on macro-micro message passing is provided, which is used to execute any of the graph contrast learning methods for improving the performance of node classification in same-different network using macro-micro message passing as described in the first aspect, and includes the following modules:

[0030] Neighborhood expansion module, used to generate k-nearest neighbor graphs and enrich the direction of message transmission between nodes;

[0031] An enhanced view generation module, used to generate enhanced views required for contrastive learning;

[0032] A view role definition module is used to determine the priority or auxiliary role of the structure view and feature view according to the homogamy and heterogamy properties of the network;

[0033] The priority view and auxiliary view alignment encoding module is used to encode and align the priority view and auxiliary view respectively, and obtain the embedded representation of the node in different views;

[0034] A contrastive training module, which is used to optimize node embedding representation using a contrastive loss function;

[0035] The node classification module is used to train and evaluate node classification tasks using the optimized node embedding representation.

[0036] In a third aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program; when the computer program is executed on a computer, the computer executes any graph comparative learning method for improving the classification performance of nodes in networks with the same or different configurations based on macro-micro message passing as described in the first aspect.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] (1) This paper proposes a new graph contrastive learning framework by integrating macro and micro message passing strategies, which significantly improves the performance of graph contrastive learning in homogeneous and heterogeneous networks and enhances the performance and generalization ability of the model in node classification tasks.

[0039] (2) At the macro level, the present invention effectively overcomes the problem of insufficient message transmission in heterogeneous networks in the prior art by adopting a superior-auxiliary view alignment coding strategy combining structural view and feature view, and enhances the model's adaptability to networks with different homogeneity levels;

[0040] (3) At the micro level, the present invention optimizes the encoder design through an adaptive self-propagation strategy, flexibly handles the role of node self-loops in feature propagation, and makes node features more detailed and accurate during the propagation process;

[0041] (4) The present invention improves the quality of node representation while reducing computational complexity, especially when processing heterogeneous networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic flow chart of a method for classifying nodes in networks with the same or different configurations based on graph comparison learning based on macro-micro message passing provided by an embodiment of the present invention;

[0043] Figure 2 Schematic diagram of an adaptive self-propagation strategy for optimizing an encoder;

[0044] Figure 3 Schematic diagram of the framework for graph contrast learning based on macro and micro messages;

[0045] Figure 4 This is a structural diagram of a network node classification system for similar and different matching based on graph comparative learning of macro-micro message passing. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0047] Example 1

[0048] like Figure 1 As shown, Embodiment 1 of the present invention provides a method for classifying nodes in networks with the same or different configurations based on graph comparison learning based on macro-micro message passing, comprising the following steps:

[0049] S1: Based on the original image Where A is the adjacency matrix, X is the node feature matrix, the cosine distance between node features is calculated, and a k-nearest neighbor graph based on the similarity of node features is constructed.

[0050] In the present invention, the cosine distance is first calculated using the node features, and the calculation formula is:

[0051]

[0052] where x i and x j The nodes v i and v j The feature vector between . After calculating the cosine distance between each node and other nodes according to this formula, k nearest neighbor nodes are selected for each node and the corresponding edges are added to construct the node feature similarity. Here, the value range of k is {5, 10, 30, 50, 70}. Usually, the original graph is a sparse graph, and the k nearest neighbor graph is a dense graph.

[0053] S2: Perform a view enhancement operation of randomly removing edges on the original graph and the k-nearest neighbor graph to generate the corresponding structural view v str and feature view v att .

[0054] In the present invention, the edge types of the original graph and the k-nearest neighbor graph are different. The edge types of the original graph are based on structural information, while the edge types of the k-nearest neighbor graph are based on node feature similarity. Generate structure view v str , and the k nearest neighbor graph generates the feature view v att Since the two views have different edge types, they belong to views of different scales. This multi-scale view enhancement operation helps graph contrastive learning capture both local structural information and global feature information, providing richer semantic expressions.

[0055] S3: According to the edge homogeneity ratio h of the original graph edge Determine the type of network configuration and then define the priority view v of the network configuration p Is the structure view v str , auxiliary view v s is the feature view v att; Priority view v of heterogeneous networks p is the feature view v att , auxiliary view v s Is the structure view v str .

[0056] In the present invention, the edge homogeneity ratio h is used edge To measure the homogeneity level of the network, the calculation formula is:

[0057]

[0058] Based on h edge We define the value of but For heterogeneous networks, if but The same network.

[0059] In the generated structure view v str and feature view v att Finally, we introduce the priority view v p and auxiliary view v s The concept of is used to clarify the different roles of these two views in the graph contrast learning process. str As the priority view, it is responsible for the main feature extraction task and dominates the propagation and learning of node features; the feature view v att As an auxiliary view, it provides supplementary information. In the heterogeneous network, the situation is the opposite. The feature view v att becomes the priority view, while the structure view v str This way of defining the roles of the priority view and the auxiliary view helps improve the generalization ability of the model and give full play to the performance advantages of graph contrast learning in homogeneous and heterogeneous networks.

[0060] S4: Use two independent ASP-SGC encoders to encode the priority view v p and auxiliary view v s Perform priority view alignment coding to obtain the node representation H after priority view coding p The node representation H after alignment encoding with the auxiliary view sa .

[0061] In this invention, the view encoding process introduces innovative improvements at both macro and micro levels. The macro level refers to the direction of message transmission, i.e., the structure view and feature view at two different scales. The micro level refers to the way of message transmission, i.e., the role played by node self-loops in the feature propagation process.

[0062] At the micro level, the present invention adopts a more flexible ASP-SGC encoder. Based on SGC, the encoder decomposes the propagation of node self-loop features into two parts: neighbor assimilation propagation and self-independent propagation through the adaptive self-propagation (ASP) strategy. The specific form is:

[0063]

[0064] The first term (1-ω)H (k) Represents the self-independent propagation part. This part of the self-loop does not act as a neighbor, but independently propagates features, emphasizing the independence of self-loop feature propagation without relying on neighbor node features. The second item represents the neighbor assimilation propagation part, which acts as a neighbor for feature propagation. Here, k represents the node of the kth layer, representing H (k) , ω is the self-loop role trade-off factor, and Θ is the trainable weight matrix.

[0065] At the macro level, the present invention adopts the method of view encoding by using the preferred view alignment encoding method. This encoding method is different from other common techniques that encode views and then directly perform contrast training. Instead, it introduces an alignment mechanism. Specifically, first use the preferred view v p and auxiliary view v s Each ASP-SGC encoder encodes the corresponding view and obtains the node representation H p and H s ; Then the ASP-SGC encoder of the auxiliary view is used to encode the priority view v from a global perspective p and compare it with the initial node representation v of the auxiliary view s Add together to generate the aligned auxiliary view and the final node representation H sa :

[0066]

[0067] Traditional graph contrastive learning relies on comparing views at the same scale, but in multi-scale scenarios, the embedding distributions of different scales are often inconsistent, resulting in poor contrastive learning results. Through the alignment encoding operation, this strategy effectively solves the problem of mismatched embedding representations in multi-scale views, ensuring that node embeddings in multi-scale views remain consistent, thereby improving the overall effect of contrastive learning in multi-scale.

[0068] S5: Node-based representation H p and H sa , define the anchor node v i The InfoNCE contrast loss function is constructed based on the positive and negative sample pairs, and the node embedding representation of the model is optimized and trained through this loss function.

[0069] In the present invention, the anchor node v is defined i The positive sample pair is Negative sample pairs include as well as The InfoNCE contrast loss is constructed as:

[0070]

[0071] where τ is the temperature parameter and s(·) is the cosine similarity. The contrast loss function maximizes the probability that the same node will appear in the priority view H. p and auxiliary view H sa The similarity of representations in the node is maximized, while minimizing the similarity of representations between different nodes, thereby effectively enhancing the distinguishability of node representations.

[0072] S6: The optimized node embedding representation is input into the logistic regression node classifier for training and evaluation.

[0073] In this paper, in order to evaluate the quality of node embedding representations generated by the graph contrastive learning method based on macro-micro message passing to improve the performance of node classification in homogeneous and heterogeneous networks, we first perform unsupervised training on the graph data (A, X) to generate node embedding representations. Subsequently, these embedding representations are input into the L2 regularized logistic regression classifier for training and evaluation of the node classification task.

[0074] To ensure the reliability of the results, five experiments were conducted on the node classification task. The classification accuracy was recorded in each experiment, and the performance of the model was evaluated by calculating the average and standard deviation of these accuracy rates.

[0075] Experimental verification:

[0076] During the experiment, seven public datasets with different homogeneity levels were used: Cora, CiteSeer, PubMed (homogeneous network datasets) and Cornell, Texas, Wisconsin, Actor (heterogeneous network datasets). All datasets adopted the fixed partition proposed in the original work.

[0077] To ensure fairness in performance comparison between different methods, all compared algorithms use the same dataset partitioning, training scheme, and classifier, and adopt consistent evaluation criteria.

[0078] The model of the present invention is compared with six supervised algorithms, namely MLP, GCN, SGC, H2GCN, UGCN and FAGCN, and six self-supervised algorithms, namely NCLA, HSAN, CSGCL, HomoGCL, GREET and ASP, in node classification on seven real data sets. The results are shown in Table 1, and the bold content indicates the best result.

[0079] Table 1 Node classification task results (%)

[0080]

[0081] Experimental results show that the proposed method significantly outperforms existing baseline methods in node classification tasks on homogeneous and heterogeneous networks. In particular, the model performs well in heterogeneous networks, demonstrating strong robustness and generalization capabilities.

[0082] Example 2

[0083] like Figure 4 As shown, based on Example 1, Example 2 of the present invention provides a network node classification system for the same and different configurations based on macro-micro message passing for graph comparative learning, including the following modules:

[0084] Neighborhood expansion module, used to generate k-nearest neighbor graphs and enrich the direction of message transmission between nodes;

[0085] An enhanced view generation module, used to generate enhanced views required for contrastive learning;

[0086] A view role definition module is used to determine the priority or auxiliary role of the structure view and feature view according to the homogamy and heterogamy properties of the network;

[0087] The priority view and auxiliary view alignment encoding module is used to encode and align the priority view and auxiliary view respectively, and obtain the embedded representation of the node in different views;

[0088] A contrastive training module, which is used to optimize node embedding representation using a contrastive loss function;

[0089] The node classification module is used to train and evaluate node classification tasks using the optimized node embedding representation.

[0090] Specifically, the system provided in this embodiment 2 is a system corresponding to the method provided in embodiment 1. Therefore, the parts in this embodiment 2 that are the same or similar to those in embodiment 1 can be referenced to each other and will not be repeated in the present invention.

[0091] Example 3

[0092] Based on Examples 1 and 2, Example 3 of the present invention provides a computer-readable storage medium, which stores a computer program; when the computer program runs on a computer, the computer executes the graph comparative learning method for improving the classification performance of nodes in the same and different matching networks based on macro-micro message passing as described in Example 1 of the present invention.

[0093] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for classifying nodes in networks with the same or different configurations based on graph contrast learning based on macro-micro message passing, characterized in that: The following steps are involved: S1: Based on the original image Where A is the adjacency matrix, X is the node feature matrix, the cosine distance between node features is calculated, and a k-nearest neighbor graph based on the similarity of node features is constructed; S2: Perform a view enhancement operation of randomly removing edges on the original graph and the k-nearest neighbor graph to generate the corresponding structural view v str and feature view v att ; S3: According to the edge homogeneity ratio h of the original graph edge Determine the type of network configuration and then define the priority view v of the network configuration p Is the structure view v str , auxiliary view v s is the feature view v att ; Priority view v of heterogeneous networks p is the feature view v att , auxiliary view v s Is the structure view v str ; S4: Use two independent ASP-SGC encoders to encode the priority view v p and auxiliary view v s Perform priority view alignment coding to obtain the node representation H after priority view coding p The node representation H after alignment encoding with the auxiliary view sa ; S5: Node-based representation H p and H sa , define the anchor node v i The InfoNCE contrast loss function is constructed based on the positive and negative sample pairs, and the node embedding representation of the model is optimized and trained through the loss function; S6: The optimized node embedding representation is input into the logistic regression node classifier for training and evaluation.

2. According to claim 1, the graph comparison learning method for improving the classification performance of nodes in the same and different configuration networks based on macro-micro message passing is characterized in that: In S1, the cosine distance calculation formula between node features is: where x i and x j The nodes v i and v j According to the cosine distance calculation result, k nearest neighbor nodes are selected for each node and corresponding edges are added to build a k-nearest neighbor graph based on node feature similarity.

3. According to claim 1, the graph comparison learning method for improving the classification performance of nodes in the same and different configuration networks based on macro-micro message passing is characterized in that: In S3, the edge homogeneity ratio h, which measures the homogeneity level of the network, edge The calculation formula is: like but For heterogeneous networks, if but The network is the same as the one provided by the user.

4. According to claim 1, the graph comparison learning method for improving the classification performance of nodes in the same and different configuration networks based on macro-micro message passing is characterized in that: In S4, the ASP-SGC encoder uses an adaptive self-propagation strategy to decompose the node self-loop feature propagation into neighbor assimilation propagation and self-independent propagation: The first term (1-ω)H (k) Represents the self-independent propagation part, which emphasizes the independence of self-loop feature propagation and does not depend on the characteristics of neighboring nodes; Item 2 represents the neighbor assimilation propagation part, in which the self-loop acts as a neighbor to propagate features; k represents the node of the kth layer, representing H (k) , ω is the weight coefficient for self-independent propagation, and Θ is the trainable weight matrix.

5. According to claim 1, the graph comparison learning method for improving the classification performance of nodes in the same and different configuration networks based on macro-micro message passing is characterized in that: In S4, the priority view v is first used p and auxiliary view v s Each ASP-SGC encoder encodes the corresponding view and obtains the node representation H p and H s ; Then the ASP-SGC encoder of the auxiliary view is used to encode the priority view v from a global perspective. p and compare it with the initial node representation v of the auxiliary view s Add together to generate the aligned auxiliary view and the final node representation H sa :

6. The graph comparison learning method for improving the classification performance of nodes in homogeneous and heterogeneous networks based on macro-micro message passing according to claim 1 is characterized in that: In S5, define the anchor node v i The positive sample pair is Negative sample pairs include as well as The InfoNCE contrast loss is constructed as: where τ is the temperature parameter and s(·) is the cosine similarity.

7. A network node classification system based on macro-micro message passing for graph contrast learning, characterized in that: The graph comparison learning method for improving the classification performance of nodes in networks with the same or different configurations based on macro-micro message transmission as described in any one of claims 1 to 6 comprises: Neighborhood expansion module, used to generate k-nearest neighbor graphs and enrich the direction of message transmission between nodes; An enhanced view generation module, used to generate enhanced views required for contrastive learning; A view role definition module is used to determine the priority or auxiliary role of the structure view and feature view according to the homogamy and heterogamy properties of the network; The priority view and auxiliary view alignment encoding module is used to encode and align the priority view and auxiliary view respectively, and obtain the embedded representation of the node in different views; A contrastive training module, which is used to optimize node embedding representation using a contrastive loss function; The node classification module is used to train and evaluate node classification tasks using the optimized node embedding representation.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is loaded and executed by a processor to implement the method for classifying nodes in a same- and different-matching network based on graph comparative learning based on macro- and micro-message passing as described in any one of claims 1 to 6.

Images