Network node automatic classification and personalized matching method based on learnable graph enhancement and enhanced comparison loss

Abstract

The invention discloses a network node automatic classification and personalized matching method based on learnable graph enhancement and enhanced comparison loss. According to the method, semantic diversity is improved through an adaptive graph enhancement module based on an attention mechanism; putting forward uniform similarity measurement of fusion structure and attribute information to construct a semantic consistent sample pair; and a double-weight modulation comparison loss function is introduced, and the comparison intensity is dynamically adjusted to relieve false negative sample interference. According to the method, the discrimination and generalization ability of graph node representation and the robustness to false negative samples are effectively improved, and the method is particularly suitable for a node classification scene with limited annotation data. And the learned embedded representation has migration capability, and can support downstream applications such as personalized matching and the like.

Description

Technical Field

[0001] The present invention relates to the field of graph representation learning, and mainly relates to a method for automatic classification and personalized matching of network nodes based on learnable graph enhancement and enhanced contrast loss. Background Art

[0002] The application of structured graph data has shown broad potential in the automatic classification and personalized matching of network nodes. For example, by constructing structures such as entity collaboration networks, attribute association graphs, and behavioral relationship graphs, we can reveal potential high-level relationship patterns between nodes, providing critical structural and semantic information support for classification and matching tasks.

[0003] However, practical applications still face two core challenges: First, mainstream graph neural network (GNN) methods rely heavily on large amounts of manual annotation, while node labels in many real-world scenarios need to be provided by experts or manually screened, resulting in high costs and poor scalability; second, existing graph representation learning methods are difficult to adapt to the semantic diversity and dynamic changes in complex graph data, and model performance and generalization capabilities are limited.

[0004] Taking node classification as an example, traditional methods often rely on manually constructed category labels. However, with the surge in data volume, manual annotation can no longer sustain large-scale system training. In personalized matching tasks, traditional collaborative filtering or content-based recommendation methods struggle to model high-order semantic dependencies between nodes, especially when faced with cross-group behavior migration, cold-start nodes, or long-tail data. Furthermore, the actual graph structure exhibits significant dynamic evolution in many applications (such as changes in user relationships, interaction log updates, and device topology adjustments), further challenging the robustness and generalization capabilities of representation learning models.

[0005] As a typical self-supervised learning paradigm, contrastive learning has been widely used in graph representation learning in recent years to alleviate label dependency. Conventional approaches include generating semantically preserved enhanced graph views through regularized strategies such as edge perturbations, feature masking, and subgraph sampling. Contrastive loss functions such as InfoNCE are then used to optimize node embedding quality by bringing positive pairs closer and negative pairs further apart. However, these general methods still face multiple limitations in complex graph data scenarios:

[0006] Inadequate adaptability of augmentation strategies to structural semantics: Fixed-rule graph augmentation strategies (such as random edge deletion and feature masking) cannot effectively address the asymmetry and sensitivity of semantic relationships in real networks. If the augmentation strength is too high, it may destroy key topological relationships; if it is too low, it may not provide sufficient positive and negative sample differentiation to achieve information gain.

[0007] Positive and negative sample pairs lack semantic consistency: Random negative sampling can easily introduce "false negatives," where nodes that are highly similar to anchor nodes in attributes or structure are mistakenly sampled as negative samples. Most methods measure similarity based solely on structure or features, ignoring the complementary expressive power of integrating the two.

[0008] Failure to utilize multi-source semantically associated group information: Current methods mostly use a single anchor pair comparison method of one positive and one negative, ignoring the semantic consistency across views or within the group, which limits the model's ability in semantic cluster modeling and diversity representation.

[0009] To address these issues, this paper proposes a method for automatic classification and personalized matching of network nodes based on learnable graph augmentation and enhanced contrastive loss. This method achieves a dynamic balance between enhanced perturbation strength and semantic preservation by introducing a learnable augmentation mechanism for structural-semantic collaborative optimization. Furthermore, it integrates a multi-source sample construction mechanism based on structural and attribute similarity to significantly enhance contrastive learning's ability to model high-order semantics and local group structures, ultimately achieving high-precision classification and matching of network nodes. Summary of the Invention

[0010] The present invention aims to provide a method for automatic classification and personalized matching of network nodes based on learnable graph enhancement and enhanced contrast loss (hereinafter referred to as ECLA). In response to the pain points of complex node semantic relationships and high manual annotation costs in large-scale structured graph data, a dynamic dual-weight modulation mechanism is proposed: when the correlation between the sample node and the anchor node is high, a larger positive contrast weight is assigned and the negative contrast weight is reduced to strengthen the intra-class compactness; otherwise, the negative weight is increased to enhance the inter-class distinction. In order to accurately measure the association strength between nodes, the method designs a hybrid measurement mechanism that integrates structural similarity and attribute similarity to achieve a comprehensive evaluation of node relationships. At the same time, the enhanced contrast loss is combined with the learnable graph enhancement strategy to adaptively generate edge weight parameters for each enhanced view while fully retaining the original topological information. By minimizing the overall contrast loss, the feature alignment effect of the positive sample pairs is improved, and the discriminative ability of the negative sample pairs is enhanced. Finally, a high-quality node embedding representation is obtained and applied to node classification and personalized matching. The present invention significantly improves the classification accuracy and personalized matching effect of network nodes, and is suitable for efficient processing and intelligent decision-making in various large-scale graph data scenarios such as academic networks, social networks, commodity co-purchasing networks, transportation networks, and Internet of Things device networks.

[0011] The present invention solves the above technical problems through the following technical solutions, which include the following steps:

[0012] Step S1: Obtain the original network dataset (such as social network, commodity co-purchase network, device interaction graph, etc.) and convert it into a graph structure representation; the graph data is defined as G = (A, X), where A∈{0,1} N×N and Represent the adjacency matrix and node feature matrix of the graph respectively, N is the number of nodes in the graph, F is the feature dimension of each node, and the node set is Each node v i Corresponding to a feature vector represents the set of real numbers;

[0013] Step S2: Use a multi-head graph attention network (GAT) to perform learnable graph enhancement. During the graph enhancement process, the original topology of the graph is retained, K enhanced graph views with adaptive topology are generated, and a graph convolutional network is used in each enhanced view to generate a different low-dimensional node representation to enhance the diversity of node embeddings.

[0014] Step S3: using a hybrid similarity measurement mechanism that integrates structural features to calculate the similarity between node pairs, which is used to evaluate the correlation and potential connection between nodes;

[0015] Step S4: Based on the enhanced contrast loss function, an enhanced contrastive learning method with a dual-weight modulation mechanism is designed. When the sample node has a high feature correlation or structural consistency with the anchor node, it is given a higher positive contrast weight and a lower negative contrast weight; conversely, its positive weight is reduced and the negative weight is increased to enhance the distinction between nodes and the aggregation within the class.

[0016] Step S5: Based on the dual-weight modulation mechanism, the contrast loss of each enhanced view is aggregated to construct an overall contrast loss function of multiple views;

[0017] Step S6: Based on the overall contrast loss, all learnable parameters of the backbone model are updated through network backpropagation, including the multi-head GAT module parameters and graph convolutional network parameters;

[0018] Step S7: After the parameters of the backbone network are learned, the node embeddings of multiple enhanced views are connected to obtain the final node embedding matrix F' is the feature dimension of node embedding;

[0019] Step S8: After obtaining the final node embedding matrix, since only some nodes in the graph have category labels, a lightweight classification module is constructed to classify the nodes, thereby achieving automatic label inference of data nodes in the network;

[0020] Step S9: Constructing a subject-node interaction matrix to generate subject embedding, fusing the subject embedding with the node embedding, establishing an association mechanism between the subject and the node, and calculating the subject's preference score for the node, thereby achieving personalized matching or recommendation;

[0021] As an embodiment of the present invention, the process of step S2 includes:

[0022] S2-1. Define an adaptive topology There are K learnable augmented views, and the node embedding matrix of the k-th view is

[0023] S2-2、v i and v j are two nodes in the graph, in the kth augmented view v i and v j The adaptive topological connections are learned in the following way:

[0024]

[0025] Among them, if A ij =0 W (k) and represents the learnable weight parameters and transformation vectors, || represents the connection operation, Representative node v i The set of neighbor nodes, LeakyReLU is the activation function;

[0026] S2-3. Define v i The embedding of the k-th view is In the graph convolutional network, the feature information of the node and its neighbors is aggregated according to the adaptive topology, and the ELU activation function is applied to learn the embedding of each node:

[0027]

[0028] Where ELU(·) represents a nonlinear activation function.

[0029] As an embodiment of the present invention, the process of step S3 includes:

[0030] S3-1. Define node v i and v j The structural similarity and feature similarity are sim G (v i ,v j ) and sim F (v i ,v j ), the mixed similarity measure is sim(v i,v j );

[0031] S3-2, Order Represents the Katz matrix, which stores the Katz index between all pairs of nodes in the graph. Based on the relationship between the number of paths between nodes and the adjacency matrix A, the calculation of the Katz index can be transformed into the following algebraic problem:

[0032]

[0033] Where I is the identity matrix, β is the weight decay factor, and the condition for the above matrix series to converge is that β is less than the inverse of the principal eigenvalue of the adjacency matrix A.

[0034] S3-3, after calculating Based on this, normalize it to calculate sim G (v i ,v j ), distributing the global connectivity score relatively evenly in the range of 0 to 1:

[0035]

[0036] Normalization represents normalization by row maximum and minimum values.

[0037] S3-4, in order to better mine node v from features i and v j The feature similarity sim F (v i ,v j ), using the classic cosine similarity:

[0038]

[0039] S3-5. Integrate the structural similarity and feature similarity of nodes into a unified similarity:

[0040] sim(v i ,v j )=(1-α)sim G (v i ,v j )+αsim F (v i ,v j )

[0041] Among them, α is a hyperparameter used to weigh the mixing ratio between the two similarities.

[0042] As an embodiment of the present invention, the process of step S4 includes:

[0043] S4-1. Assuming that the node has a potential category, define the node v i and node v j The latent class is g(v i ) and g(v j ), and then model the positive and negative sample distribution based on the node potential category.

[0044] S4-2. For anchor node v i , define the nodes v that are observed to share the same latent class with the anchor node j The probability of It is expressed as follows:

[0045]

[0046] Similarly, for the anchor node v i , define the node v that is observed not to have the same latent class as the anchor node j The probability of It is expressed as follows:

[0047]

[0048] S4-3. According to Bayes' theorem, we have:

[0049]

[0050] Where p(v j ) is uniformly distributed, is the sample node v i and anchor node v j The probability of sharing the same latent class, is the sample node v i and anchor node v j The probability of not having the same latent class.

[0051] S4-4、Given sim(v i ,v j ) to measure node v i and node v j The similarity of With sim(v i ,v j ) positively correlated, With sim(v i ,v j ) is negatively correlated, that is:

[0052]

[0053] in and They are monotonically increasing transformation and monotonically decreasing transformation respectively.

[0054] S4-5. Based on S4-3 and S4-4, it can be intuitively shown that similar nodes are more likely to be sampled as positive samples, and dissimilar nodes are more likely to be sampled as negative samples:

[0055]

[0056] S4-6. In order to introduce more positive samples into the contrast loss and reduce the interference of false negative samples, the enhanced contrast loss with dual weight modulation is introduced. It is expressed as follows:

[0057]

[0058] where f(·) represents the function that generates low-dimensional embeddings, τ is the temperature parameter for contrastive learning, and Represent the positive sample set and the negative sample set respectively, v' i is the anchor node v i In other view enhancements, weight coefficients and Represents node v j As the probability weights of positive and negative samples.

[0059] S4-7, using the exponential function with temperature parameter and Transformation modeling, flexibly adjusting between conservative and aggressive estimates based on different temperature parameters and as follows:

[0060]

[0061] where τ p and τ n is the temperature parameter.

[0062] As an embodiment of the present invention, the process of step S5 includes:

[0063] S5-1. Define view Figure 1 Heshi Figure 2 Generated node v i The embeddings are and In sight Figure 1 Heshi Figure 2 Calculated in Related enhanced contrast loss The definition is as follows:

[0064]

[0065] Among them, pos represents the contribution of positive samples to the loss function, and neg represents the contribution of negative samples to the loss function;

[0066] S5-2, Vision Figure 1 Heshi Figure 2 The embedding matrix composed of all node embeddings is represented as H (1) and H (2) , the enhanced contrast loss between two views is defined as the average of all node contrast losses, which can be expressed as:

[0067]

[0068] S5-3, the method generates the embedding matrix of K views , the total contrast loss target Defined as anchor view H (l) The average of the sum of contrast losses between and all other enhanced views is expressed as follows:

[0069]

[0070] As an embodiment of the present invention, the step S8 process includes:

[0071] S8-1. Input the node embedding matrix with labels Define the weight matrix N T is the number of nodes with labels, and C is the number of categories in the sample. The node embedding matrix is ​​augmented, and the augmented node embedding matrix is in The length is N, which is all 1 T Column vector of .

[0072] S8-2. Perform linear combination on the node embedding matrix, expressed as follows:

[0073]

[0074] in Every element Z jc Represents node v j The raw score belongs to category c.

[0075] S8-3, map the linear output Z to the (0,1) interval through the Sigmoid function σ(.), and define Among them, P jc For sample v j The probability of belonging to category c is expressed as follows:

[0076] P=σ(Z)

[0077]

[0078] S8-4, input node label matrix using one-hot encoding Each element of Y jc Represents node v j Whether it belongs to category c, construct the cross entropy loss function Right now:

[0079]

[0080] S8-5. To prevent overfitting, add regularization loss Right now:

[0081]

[0082] where λ is used to control the regularization strength, ||·|| F is the Frobenius norm of the matrix.

[0083] S8-6. Constructing the total loss function Update the classifier parameters:

[0084]

[0085] S8-7. Calculate the category probabilities of all unlabeled nodes and select the category with the highest probability as the classification result.

[0086] As an embodiment of the present invention, the process of step S9 includes:

[0087] S9-1. In the personalized matching system, there are M subject entities (such as users, devices, etc.) and N resource nodes (nodes in the network dataset), and a binary interaction matrix R∈{0,1} is constructed. M×N , R ui =1 represents the subject u to node v i There is interaction (such as click, collection, access, control, etc.), otherwise R ui =0;

[0088] S9-2. For each subject entity u, average the embedding representations of all its interacted nodes to obtain the user vector e u , which is expressed as follows:

[0089]

[0090] in is the set of interactions of subject u, is node v i The node embedding of comes from the embedding matrix H in step S7;

[0091] S9-3, in order to estimate the subject u's response to node v i The preference degree of e u With h i Concatenate and input into a multi-layer perceptron for scoring:

[0092]

[0093] Among them [e u ;h i ] represents the concatenation operation of two vectors. MLP is a fully connected neural network with a nonlinear activation function. The output Represents the subject u to node v i preference score;

[0094] S9-4. Use the stationary ranking loss function to train the recommendation model. The loss function is expressed as follows:

[0095]

[0096] It is a training set consisting of all known interaction samples;

[0097] S9-5. For each subject entity u, calculate its prediction score for all candidate nodes Sort by scores from high to low, and select the top several nodes as the final personalized matching results.

[0098] Beneficial effects:

[0099] Aiming at two types of tasks, automatic node classification and personalized matching in structured graph data, the present invention proposes a graph representation learning method that integrates a learnable graph enhancement strategy, a node hybrid similarity measurement mechanism, and an enhanced contrast loss with dual weight modulation. This method effectively alleviates the problem of false negative sample interference in contrastive learning, and improves the model's ability to express positive sample relationships and the overall representation quality. In the node classification task, the method of the present invention has been verified in comparative experiments on three public graph datasets and 13 mainstream algorithms, showing performance superior to existing graph representation learning methods, proving that it has good discrimination ability and generalization effects under complex graph structures. At the same time, the proposed graph enhancement and contrast mechanism has good versatility and can also be applied to graph task scenarios such as personalized matching, and has broad promotion potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] Figure 1 It is a schematic diagram of the system architecture of the present invention.

[0101] Figure 2 This is a comparative analysis chart of the learning temperature hyperparameter experiment.

[0102] Figure 3 This is a similarity trade-off coefficient hyperparameter experimental analysis diagram. DETAILED DESCRIPTION

[0103] The present invention will be further described below with reference to the accompanying drawings. It should be understood that the embodiments described herein are only a portion of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0104] Research work related to this invention:

[0105] Graph Enhancement Methods

[0106] When processing graph data, challenges such as structural sparsity, noise sensitivity, topological complexity, and specific task requirements such as generative modeling and contrastive learning are common. To address these issues and improve model performance, graph augmentation techniques have become an effective means of expanding data diversity or optimizing the model training process. In graph generative models, some methods (such as EDP-GNN) construct enhanced views by adding perturbed noise to the graph structure and use this as a basis to learn the latent distribution of the original graph data. In graph contrastive learning tasks, the goal of graph augmentation is to generate positive examples that both preserve the original semantic information and are structurally compatible with the original graph. Methods for graph augmentation can be roughly divided into rule-driven and learning-driven approaches. Within rule-driven graph contrastive learning methods, different augmentation strategies reflect their respective design concepts. For example, DGI achieves view enhancement by shuffling the node attribute matrix row by row; GRACE uses random edge removal and node feature masking to simultaneously perturb structure and attributes; GCA introduces adaptive edge deletion and feature masking mechanisms to dynamically identify important connection relationships and key feature dimensions; GraphCL and GCC generate subgraphs based on random walks to construct semantically consistent comparison samples; SUBGCON uses the neighborhood structure of the central node for local enhancement, and MVGRL combines graph diffusion technology to construct local and global structural views to capture multi-scale information. At the same time, some learning-driven graph enhancement methods also show unique advantages. For example, gCooL uses graph community partitioning to integrate clustering structure with representation learning; NCLA generates fully learnable graph enhancement views based on the multi-head graph attention mechanism, while maintaining the topological invariance of the overall structure.

[0107] Comparative Learning Objectives

[0108] The setting of graph comparison objectives is closely related to the comparison paradigm employed. Currently, graph comparison learning primarily employs two core models: node-graph comparison and node-node comparison. In the node-graph comparison paradigm, typical methods (such as DGI) achieve alignment between node representations and the overall graph structure by comparing the representations of nodes in the corrupted view with those in the original graph. During training, neighbor sampling strategies are often incorporated to improve efficiency. In particular, during the negative sample generation phase, structural perturbations are strategically constructed by sampling from the neighborhood of anchor nodes. In the node-node paradigm, node-level representations in the two views are directly compared. The InfoNCE loss function is often constructed within the noise contrastive estimation (NCE) framework, defining enhanced view pairs from the same node as positive samples and combinations of different nodes as negative samples. However, previous studies have found that this negative sample sampling mechanism is prone to "sampling bias." Specifically, the "difficult" negative samples selected are often semantically highly similar to the anchor nodes, constituting false negatives. This, in turn, forces nodes that should be clustered to be pulled apart, compromising the model's expressive power. To address the limitations of negative sample handling in the traditional InfoNCE framework, subsequent research has proposed improved strategies from various perspectives. For example, BGRL completely avoids the need to define negative samples by introducing a delayed version of the encoder to guide representation learning; ProGCL optimizes negative sampling methods to improve sample quality. Some research has further proposed idealized contrastive objectives, which include more positive sample pairs in the optimization objective and explicitly exclude potential false negative samples. Based on this, more practical weak contrastive objective designs have been proposed, balancing sample selection quality and computational feasibility.

[0109] The present invention provides a graph contrast representation learning method based on learnable graph data enhancement and dual-weighted modulated enhanced contrast loss, combined with Figure 1 , including the following steps:

[0110] Step S1: Obtain the original network dataset (such as social network, commodity co-purchase network, device interaction graph, etc.) and convert it into a graph structure representation. We consider a graph data G = (A, X), where A∈{0,1} N×N and Represent the adjacency matrix and node feature matrix of the graph respectively, N is the number of nodes in the graph, F is the feature dimension of each node, and the node set is Each node v i Corresponding to a feature vector Define a matrix To represent the embedding of the graph nodes finally learned by the method, where F'<<F.

[0111] Step S2: Use a multi-head GAT to perform learnable graph enhancement. During the graph enhancement process, the original topology of the graph is retained, K enhanced graph views with adaptive topology are generated, and a graph convolutional network is used in each enhanced view to generate a low-dimensional node representation with differences. A GAT-based learnable graph enhancement method is adopted. In a GAT with K attention heads, K learnable enhanced views with adaptive topology structures are generated, denoted as In the kth augmented view, node v i and node v j The adaptive topological connection relationship between them will be learned in the following way:

[0112]

[0113] Among them, if A ij =0 W (k) and represents the learnable weight parameters and transformation vectors, || represents the connection operation, Representative node v i The set of neighbor nodes, LeakyReLU is the activation function.

[0114] Define v i The embedding of the k-th view is The graph encoder aggregates the features of a node and its neighbors according to the adaptive topology and then applies the ELU activation function to learn an embedding representation for each node:

[0115]

[0116] Among them, ELU(·) represents the nonlinear activation function and the node embedding matrix of the k-th view is

[0117] Step S3: Given a node v i and node v j The structural similarity between G (v i ,v j ) and feature similarity sim F (v i ,v j ), design a fusion similarity sim(v i ,v j ), which combines these metrics into a unified similarity score, thus capturing complementary information from both.

[0118] The calculation of network proximity is a widely used and effective method in data analysis applications, including link prediction and collaborative filtering. The Katz index is often used to calculate network proximity and has been shown to be one of the best performing path-based link prediction algorithms. Node proximity focuses more on measuring the proximity between nodes in the network topology, so it is reasonable to use the Katz index to reflect the structural similarity of nodes. Let Represents the Katz matrix, which stores the Katz index between all pairs of nodes in the graph. Based on the relationship between the number of paths between nodes and the adjacency matrix A, the calculation of the Katz matrix can be transformed into the following algebraic problem:

[0119]

[0120] Where I is the identity matrix, β is the weight decay factor, and the condition for the above matrix series to converge is that β is less than the inverse of the main eigenvalue of the adjacency matrix A. Although the matrix element values ​​are restricted to the range of 0 to 1, most values ​​tend to be concentrated in the area close to 0. The main reason for this is that when calculating the Katz index, in order to increase the importance of short paths, the damping factor α is set to a sufficiently small value. Even if long paths exist, their contribution to the total is very small. Min-max normalization is applied row-wise, which makes the global connectivity score relatively evenly distributed in the range of 0 to 1:

[0121]

[0122] In order to mine the feature similarity between nodes from the features F (v i ,v j ), the present invention adopts the classic cosine similarity. The feature vectors representing node features in graph data are usually highly sparse, but cosine similarity can still provide meaningful results in this case because it mainly depends on the position and value of non-zero elements, that is:

[0123]

[0124] The structural similarity and feature similarity of the integrated nodes are converted into a unified similarity. α is a hyperparameter used to weigh the mixing ratio between the two similarities. The integration process is as follows:

[0125] sim(v i ,v j )=(1-α)sim G (v i ,v j )+αsim F (v i ,v j)

[0126] When evaluating node similarity, relying solely on structural or feature information can overlook key factors. For example, in a social network, two users may have similar interests (feature similarity) but completely different friend circles (structural dissimilarity). Conversely, users with structural similarity (e.g., overlapping communities) may lack feature similarity (e.g., differing interests). Combining these two dimensions can improve the accuracy of identifying truly similar nodes.

[0127] Step S4: Even if the true labels of the nodes in the network are available, in self-supervised learning, it is impossible to obtain the downstream semantic categories, let alone the true labels. Here, it is assumed that there is a set of discrete potential categories Used to represent the true semantics of each node.

[0128] set up is a function that assigns a latent class to a node. Next, we Two sampling distributions based on anchor nodes are constructed above. For an anchor node v i ,use It is observed that any node v shares the same latent class with the anchor node j The probability of . Similarly, It means that any node v is observed j With anchor node v i with the probabilities of different latent classes. Note that, and The subscripts in the table indicate that they are for the anchor node v i Specific.

[0129] According to Bayes' theorem, we have:

[0130]

[0131] Where p(v i ) is a uniform distribution across all nodes in the network. is the sample node v i and anchor node v j The probability of belonging to the same latent class, is the sample node v i and anchor node v j The probability of not having the same latent class.

[0132] If the node v i and node v j The more similar they are, the more likely they are to belong to the same latent class, which is intuitive. Given a function to measure v i and v j The similarity of With sim(v i ,v j ) positively correlated, With sim(v i ,v j ) is negatively correlated, that is:

[0133]

[0134] in and They are monotonically increasing transformation and monotonically decreasing transformation, which intuitively expresses that samples that are more similar to v are more likely to be sampled as positive samples, and samples that are less similar are more likely to be sampled as negative samples, that is:

[0135]

[0136] In order to introduce more positive samples into the contrast loss and reduce the interference of false negative samples, the enhanced contrast loss with dual weight modulation is introduced. It is expressed as follows:

[0137]

[0138] where f(·) represents the function that generates low-dimensional embeddings, τ is the temperature parameter for contrastive learning, and Represent the positive sample set and the negative sample set respectively, v' i is the anchor node v i In other view enhancements, weight coefficients and Represents node v j As the probability weights of positive and negative samples.

[0139] Further, let And use the exponential function with temperature parameter to and Transformation estimation, flexible adjustment between conservative and aggressive estimation based on different temperature parameters and as follows:

[0140]

[0141] where τ p and τ n is a temperature parameter that affects the selection probability of high-similarity and low-similarity samples.

[0142] Step S5, define the view Figure 1 Heshi Figure 2 Generated node v i The expressions are and Use self-view Figure 1 Embed As anchors, positive contrast signals come from three aspects: cross-view node alignment (anchors and from the perspective of Figure 2 The same node to form positive sample pairs), intra-view cross-node similarity (weighted positive comparison between anchor nodes and semantically similar nodes), and cross-view cross-node similarity (weighted positive comparison between anchor nodes and view nodes). Figure 2 The negative contrast signal comes from two aspects: negative sampling within the view (the weighted positive contrast between the anchor node and the view). Figure 1 weighted negative contrast between semantically dissimilar nodes in the Figure 2 weighted negative contrast between semantically dissimilar nodes in ). Similar to Here, the semantically similar and semantically dissimilar node sets of anchor nodes are set to The strength of the distinction between positive and negative samples is controlled by the weight coefficient. Figure 1 Heshi Figure 2 Computing and Related enhanced contrast loss definition as follows:

[0143]

[0144] Among them, pos represents the contribution of positive samples to the loss function, neg represents the contribution of negative samples to the loss function, and τ is the temperature hyperparameter of graph contrastive learning. Considering the symmetry between the two views, when selecting the view Figure 2 Embed When embedded as an anchor point, the enhanced contrast loss can be defined symmetrically See Figure 1 Heshi Figure 2 The final enhanced contrast loss is defined as the average over all nodes:

[0145]

[0146] This method generates K semantically preserved enhanced views and generates the embedding matrix of K views When K>2, the node embedding matrix H in the lth view is randomly selected. (l) As the anchor view. The total contrast loss target It is defined as the average of the contrast loss between the anchor view and all other enhanced views, that is:

[0147]

[0148] Step S6, minimize the total contrast loss target, and the network backpropagation updates the parameters of the backbone network. The feature alignment between positive sample pairs is systematically enhanced, while the discriminative separation between negative sample pairs is amplified. Specifically, this optimization promotes the formation of a semantically driven geometric structure of node embeddings in the representation space: for any anchor node, its embedding vector produces a local clustering effect with the embeddings of semantically similar nodes in the view and across views, while establishing a global repulsive relationship with the embeddings of different nodes. By adjusting the gradient propagation mechanism of the encoder parameters, the compact clustering of similar node embeddings related to downstream tasks in the Euclidean space is ensured, while maintaining distinguishable marginal intervals between different node distributions across downstream applications.

[0149] Step S7: After the parameters of the backbone network are learned, the node embeddings of multiple enhanced views are connected to obtain the final node embedding matrix H.

[0150] Step S8: Input the node embedding matrix with labels Define the weight matrix N T is the number of nodes with labels, and C is the number of categories in the sample. The node embedding matrix is ​​augmented, and the augmented node embedding matrix is in The length is N, which is all 1 T A linear combination of the node embedding matrices is performed, which is expressed as follows:

[0151]

[0152] in Every element Z jc represents the original score that sample j belongs to category c.

[0153] Map the linear output Z to the (0,1) interval through the Sigmoid function σ(.), and define Among them, P jc is the probability that sample j belongs to category c, expressed as follows:

[0154] P=σ(Z)

[0155]

[0156] Input node label matrix using one-hot encoding Constructing the cross entropy loss function And add regularization loss Right now:

[0157]

[0158] where λ is used to control the regularization strength, ||·|| F is the Frobenius norm of the matrix.

[0159] Constructing the total loss function Update the classifier parameters:

[0160]

[0161] Calculate the category probabilities of all unlabeled nodes and select the category with the highest probability as the classification result.

[0162] Step S9: In the personalized matching system, there are M subject entities (such as users, devices, etc.) and N resource nodes (nodes in the network dataset), and a binary interaction matrix R∈{0,1} is constructed. M×N , R ui =1 represents the subject u to node v i There is interaction (such as click, collection, access, control, etc.), otherwise R ui =0.

[0163] For each subject entity u, the embedding representations of all its interacted nodes are averaged to obtain the user vector e u , which is expressed as follows:

[0164]

[0165] in is the set of interactions of subject u, is node v i The node embeddings come from the embedding matrix H in step S7.

[0166] In order to estimate the effect of subject u on node v i The preference degree of e u With h i Concatenate and input into a multi-layer perceptron for scoring:

[0167]

[0168] Among them [e u ;h i ] represents the concatenation operation of two vectors. MLP is a fully connected neural network with a nonlinear activation function. The output Represents the subject u to node v i preference score;

[0169] The stationary ranking loss function is used to train the recommendation model. The loss function is expressed as follows:

[0170]

[0171] It is a training set consisting of all known interaction samples;

[0172] For each subject entity u, calculate its prediction score for all candidate nodes Sort by scores from high to low, and select the top several nodes as the final personalized matching results.

[0173] Experimental analysis

[0174] In order to evaluate the effectiveness of the multi-behavior recommendation model proposed in this paper, a series of experiments were designed to address the following research questions:

[0175] Experimental data

[0176] The effectiveness of the model of the present invention is evaluated on three public datasets: Cora, CiteSeer and Coauthor CS. Table 1 shows the basic information of the three datasets.

[0177] Table 1 Experimental dataset

[0178]

[0179] Baseline Model

[0180] For a more comprehensive evaluation, the performance of our proposed method in semi-supervised node classification is compared with several representative graph neural network methods:

[0181] Semi-supervised graph neural network methods

[0182] GCN: Updates the feature representation of each node by aggregating information from neighboring nodes.

[0183] GAT: By assigning different weights to each node's neighbor nodes, neighbor information can be aggregated more flexibly.

[0184] Semi-supervised graph contrastive learning method

[0185] CGPN: We propose a novel graph neural network designed to efficiently propagate small amounts of label information throughout the entire graph structure.

[0186] CG3: To introduce contrastive learning into semi-supervised graph learning and fully utilize the information in unlabeled data by designing a semi-supervised contrastive loss objective.

[0187] Self-supervised graph contrastive learning method

[0188] DGI relies on maximizing the mutual information between the graph augmented representation and the currently extracted graph information, both of which are generated through the established graph convolutional network architecture.

[0189] GMI: directly compares the input (i.e., the subgraph consisting of the input neighborhood) and output (i.e., the hidden representation of each node) of the encoder to derive mutual information.

[0190] MVGRL: Proposes to maximize the mutual information between node-level and graph-level representations across multiple views.

[0191] GRACE: It considers corruption at both the topology and node attribute levels to generate two views and learns node representations by maximizing the consistency of node representations in the two views.

[0192] GCA: focuses on preserving important connection structures during graph enhancement while destroying relatively unimportant node features.

[0193] SUGRL: Constructs multiple loss functions to explore the complementary information between structural information and neighborhood information to expand the difference between categories, and reduces the variation within categories by adding an upper limit loss, thereby achieving a limited distance between the positive sample embedding and the anchor embedding.

[0194] AFGRL: Instead of designing a specific graph enhancement method, it constructs positive and negative samples for contrastive learning by mining node pairs that share local structural information and global semantics in the graph, thereby achieving contrastive learning.

[0195] NCLA: By introducing multiple positive samples in the contrastive loss, the network topology is explicitly used as a supervisory signal, thereby leveraging the structural relationships within the graph to guide the learning of discriminative node embeddings.

[0196] Performance Comparison

[0197] The node embeddings are learned in a self-supervised manner and then used to train and test a logistic regression (LR) classifier with L2 regularization to achieve semi-supervised node classification. We first studied the case of extreme lack of node labels, and the number of training nodes c for each category was selected in the range of {1, 2, 3, 4}. In this environment with extreme lack of node labels, no validation set with additional labels was used for model selection. Experiments were then conducted under relatively sufficient label conditions. For Cora and CiteSeer, we randomly selected 20 nodes as training sets, 500 nodes as validation sets, and the rest as test sets. For Coauthor CS, 20 nodes per category were randomly selected as training sets, 30 nodes per category were randomly selected as validation sets, and the rest were used as test nodes. The three benchmark datasets were randomly divided into training sets, validation sets, and test sets 20 times each, and the average performance of all methods under the same division method was recorded. The following table shows the classification performance of the proposed method and the baseline on the three datasets.

[0198] Table 2 Comparison of experimental performance of the present invention on three datasets.

[0199]

[0200]

[0201] Under extremely sparse label supervision (1-4 labeled samples per class), ECLA demonstrates significant performance improvements over state-of-the-art baselines on all three benchmark datasets. ECLA's superior performance can be attributed to two aspects. First, compared to traditional contrastive learning methods that rely on hand-crafted data augmentation strategies, ECLA adaptively learns edge augmentation coefficients from the original graph structure. This approach effectively avoids the topological distortion that traditional methods can cause while minimizing the negative impact on downstream tasks. In the context of extremely scarce node labels, fully leveraging the intrinsic structure and semantic information in the network becomes particularly important. ECLA's learnable augmentation mechanism more accurately captures the essential structural features of the graph while maintaining semantic consistency. By dynamically evaluating the importance of neighboring nodes, ECLA generates semantically preserved augmented views, thus providing more reliable representation learning for downstream classification tasks. Second, the proposed dual-weight modulated augmentation loss systematically integrates graph structural similarity and feature space similarity to determine the contrast weights for positive and negative sample pairs. This sophisticated weighting scheme explicitly considers the potential semantic relevance between node pairs through a multi-view similarity metric. Specifically, graph similarity assesses topological proximity, while feature similarity captures consistency at the attribute level, forming a comprehensive semantic-aware contrastive learning framework. This design effectively alleviates the false negative problem while strategically enriching the positive sample set with semantically similar nodes. More importantly, this mechanism promotes knowledge propagation between semantically consistent nodes by optimizing gradient distribution. The resulting neighborhood aggregation process significantly improves embedding quality by preserving semantic homogeneity in the latent representation.

[0202] Ablation experiments

[0203] In order to systematically evaluate the effectiveness of the core components of the present invention, the independent effects of the positive and negative weight adjustment mechanisms are explored through ablation experiments. In the dual weight modulation enhancement loss, the positive weight coefficient is used to introduce more positive samples, while the negative weight coefficient is intended to reduce the interference effect of false negative samples. To this end, we constructed three progressive variants for verification: ECLA-Pos only retains the positive weight adjustment mechanism and adopts equal weight treatment for negative samples within and between views to verify the necessity of the negative sample weighting strategy; ECLA-Neg only retains the negative weight adjustment mechanism and strictly adopts the cross-view node self-alignment mode (i.e., only compares the embeddings of the same node in different views) for positive sample comparison, avoiding the introduction of additional positive samples, thereby decoupling the contribution of the positive sample expansion strategy; as a baseline control, ECLA-Plain completely removes the positive and negative weight mechanism and returns to the classic multi-view InfoNCE framework, where the node embedding of one view is used as the anchor, the corresponding node embedding of the other view is used as the positive sample, and all other cross-view and intra-view embeddings are used as negative samples.

[0204] The performance of ECLA and its variants on the node classification task is systematically summarized in Table 2. Our empirical analysis reveals three key observations: First, compared to the full ECLA framework, ECLA-Pos and ECLA-Neg show performance degradation on all three benchmark datasets, with ECLA-Neg's performance degradation being particularly significant. Second, ECLA-Plain further deteriorates, slightly outperforming ECLA-Neg. Third, the most significant performance gap occurs on the CiteSeer and Coauthor CS datasets, where ECLA-Neg suffers a 21.1% and 24.6% drop in accuracy compared to ECLA, respectively, highlighting the profound impact of positive weight coefficients on representation learning on these datasets. These experimental results ultimately demonstrate that both positive and negative weights are indispensable for enhancing contrastive objectives, and they jointly drive improvements in node representation learning in a complementary manner.

[0205] Table 3. Node classification results of ECLA and its variants on three datasets.

[0206]

[0207] Hyperparameter Experiment Analysis

[0208] Adjust the value of the contrastive learning temperature hyperparameter τ from 1 to 10 and report the results on Figure 2 Experimental results show that the ECLA model achieves better performance when τ is smaller. The temperature parameter τ is used to adjust the sharpness of the similarity distribution. Larger values ​​of τ will lead to over-smoothing of the similarity differences between samples.

[0209] Regarding the impact of the similarity trade-off coefficient α, three magnitude ranges were systematically tested in the study: fine-grained intervals (0.001-0.009), medium-scale values ​​(0.01-0.09), and coarse-grained settings (0.1-0.9), e.g. Figure 3 As shown, the performance evolution at different α scales exhibits a dataset-dependent pattern.

[0210] Summarize

[0211] This paper proposes a network node automatic classification and personalized matching method (ECLA) based on learnable graph enhancement and enhanced contrast loss, aiming to improve the representation quality and task adaptability of graph structure data in low-label or complex semantic scenarios. The method integrates three key innovations in its overall architecture: first, it constructs a learnable graph enhancement module based on the attention mechanism, which can adaptively adjust the edge weights according to the structural and semantic relationships between nodes, while preserving the topological structure pattern and enhancing the view diversity, avoiding the structural destruction and semantic loss caused by traditional rule enhancement strategies; second, it designs a unified hybrid similarity measurement method, which integrates graph structure similarity and node feature similarity to construct more discriminative positive and negative sample pairs; third, it introduces a dual-weight modulated enhanced contrast loss function, which dynamically balances the weights of positive and negative contrast items through a distribution-aware soft sampling strategy, effectively alleviating the interference of false negative samples and improving the clustering and generalization ability of the model representation.

[0212] In experimental verification, ECLA performed well in node classification tasks on multiple public graph datasets, especially when the number of labeled samples was extremely limited, with significantly better accuracy than existing mainstream graph comparative learning methods and semi-supervised graph neural network models. At the same time, the learned node representations have good versatility and can be applied to downstream graph structure tasks such as personalized matching, showing strong promotion potential and application value. The above is only a preferred embodiment of the present invention. Without departing from the basic principles of the present invention, any structural adjustments and module replacements should be considered within the scope of protection of the present invention.

Claims

1. An automatic classification and personalized matching of network nodes based on learnable graph augmentation and enhanced contrast loss, characterized by: The following steps are involved: Step S1: Obtain the original network dataset (such as social network, commodity co-purchase network, device interaction graph, etc.) and convert it into a graph structure representation; the graph data is defined as G = (A, X), where A∈{0,1} N×N and Represent the adjacency matrix and node feature matrix of the graph respectively, N is the number of nodes in the graph, F is the feature dimension of each node, and the node set is Each node v i Corresponding to a feature vector represents the set of real numbers; Step S2: Use a multi-head graph attention network (GAT) to perform learnable graph enhancement. During the graph enhancement process, the original topology of the graph is retained, K enhanced graph views with adaptive topology are generated, and a graph convolutional network is used in each enhanced view to generate a different low-dimensional node representation to enhance the diversity of node embeddings. Step S3: Using a hybrid similarity measurement mechanism that integrates structural features, the similarity between node pairs is calculated to evaluate the correlation and potential connection between nodes. Step S4: Based on the enhanced contrast loss function, an enhanced contrast learning method with a dual-weight modulation mechanism is designed. When the sample node has a high feature correlation or structural consistency with the anchor node, it is given a higher positive contrast weight and a lower negative contrast weight; On the contrary, its positive weight is reduced and its negative weight is increased to enhance the distinction between nodes and the aggregation within the class; Step S5: Based on the dual-weight modulation mechanism, the contrast loss of each enhanced view is aggregated to construct an overall contrast loss function of multiple views; Step S6: Based on the overall contrast loss, all learnable parameters of the backbone model are updated through network backpropagation, including the multi-head GAT module parameters and graph convolutional network parameters; Step S7: After the parameters of the backbone network are learned, the node embeddings of multiple enhanced views are connected to obtain the final node embedding matrix F' is the feature dimension of node embedding; Step S8: After obtaining the final node embedding matrix, since only some nodes in the graph have category labels, a lightweight classification module is constructed to classify the nodes, thereby achieving automatic label inference of data nodes in the network; Step S9: Construct a subject-node interaction matrix to generate subject embedding, fuse the subject embedding with the node embedding, establish an association mechanism between the subject and the node, and calculate the subject's preference score for the node, thereby achieving personalized matching or recommendation.

2. The method according to claim 1, characterized in that The step S2 includes the following sub-steps: S2-1. Define an adaptive topology There are K learnable augmented views, and the node embedding matrix of the k-th view is S2-2、v i and v j are two nodes in the graph, in the kth augmented view v i and v j The adaptive topological connections are learned in the following way: Among them, if A ij =0 W (k) and represents the learnable weight parameters and transformation vectors, || represents the connection operation, Representative node v i The set of neighbor nodes, LeakyReLU is the activation function; S2-3. Define v i The embedding of the k-th view is In the graph convolutional network, the feature information of the node and its neighbors is aggregated according to the adaptive topology, and the ELU activation function is applied to learn the embedding of each node: Where ELU(·) represents a nonlinear activation function.

3. The method according to claim 1, characterized in that The step S3 includes the following sub-steps: S3-1. Define node v i and v j The structural similarity and feature similarity are sim G (v i ,v j ) and sim F (v i ,v j ), the mixed similarity measure is sim(v i ,v j ); S3-2, Order represents the Katz matrix, which stores the Katz index between all pairs of nodes in the graph. According to the relationship between the number of paths between nodes and the adjacency matrix A, the calculation of the Katz index is transformed into the following algebraic problem: Where I is the identity matrix, β is the weight decay factor, and the condition for the above matrix series to converge is that β is less than the inverse of the principal eigenvalue of the adjacency matrix A; S3-3, after calculating Based on this, normalize it to calculate sim G (v i ,v j ), distributing the global connectivity score relatively evenly in the range of 0 to 1: Normalization represents the normalization of the maximum and minimum values ​​of the rows; S3-4, in order to better mine node v from features i and v j The feature similarity sim F (v i ,v j ), using the classic cosine similarity: S3-5. Integrate the structural similarity and feature similarity of nodes into a unified similarity: sim(v i ,v j )=(1-α)sim G (in i ,v j )+αsim F (in i ,v j ) Among them, α is a hyperparameter used to weigh the mixing ratio between the two similarities.

4. The method according to claim 1, wherein The step S4 includes the following sub-steps: S4-1. Assuming that there is a potential classification of nodes, define node v i and node v j The potential classes are g(v i ) and g(v j ), and accordingly establish a consistency relationship based on the node potential category for distribution modeling of positive and negative samples; S4-2. For anchor node v i , define the node v that is observed to have the same latent class as the anchor node j The probability of It is expressed as follows: Similarly, for the anchor node v i , define the node v that is observed to have a different latent class from the anchor node j The probability of It is expressed as follows: S4-3. According to Bayes' theorem, we have: Where p(v j ) is uniformly distributed, is the sample node v i and anchor node v j Sharing the same latent class is the sample node v i and anchor node v j The probability of not having the same latent class. S4-4、Given sim(v i ,v j ) to measure node v i and node v j The similarity of With sim(v i ,v j ) positively correlated, With sim(v i ,v j ) is negatively correlated, that is: in and They are monotonically increasing transformation and monotonically decreasing transformation respectively; S4-5. Based on S4-3 and S4-4, it can be intuitively shown that similar nodes are more likely to be sampled as positive samples, and dissimilar nodes are more likely to be sampled as negative samples: S4-6. In order to introduce more positive samples into the contrast loss and reduce the interference of false negative samples, the enhanced contrast loss with dual weight modulation is introduced. It is expressed as follows: where τ is the temperature parameter for contrastive learning, and Represent the positive sample set and the negative sample set respectively, v' i is the anchor node v i In other view enhancements, weight coefficients and Represents node v j As the probability weight of positive and negative samples, f(·) represents the function that generates low-dimensional embedding; S4-7, using the exponential function with temperature parameter and Transformation modeling, flexibly adjusting between conservative and aggressive estimates based on different temperature parameters and as follows: where τ p and τ n is the temperature parameter.

5. The method according to claim 1, wherein The step S5 includes the following sub-steps: S5-1. Define the node v generated by view 1 and view 2 i The embeddings are and The sum calculated in View 1 and View 2 Related enhanced contrast loss The definition is as follows: Among them, pos represents the contribution of positive samples to the loss function, and neg represents the contribution of negative samples to the loss function; The embedding matrix composed of all node embeddings of S5-2, view 1 and view 2 is expressed as H (1) and H (2) , the enhanced contrast loss between two views is defined as the average of all node contrast losses, expressed as: S5-3, the method generates the embedding matrix of K views Total contrast loss objective Defined as anchor view H (l) The average of the sum of contrast losses between and all other enhanced views is expressed as follows:

6. The method according to claim 1, characterized in that The step S8 includes the following sub-steps: S8-1. Input the node embedding matrix with labels Define the weight matrix N T is the number of nodes with labels, C is the total number of possible categories in the sample; the node embedding matrix is ​​augmented, and the augmented node embedding matrix in The length is N, which is all 1 T Column vector of ; S8-2. Perform linear combination on the node embedding matrix, expressed as follows: in Every element Z jc Represents node v j The original score belongs to category c; S8-3, map the linear output Z to the (0,1) interval through the Sigmoid function σ(.), and define Among them, P jc For node v j The probability of belonging to category c is expressed as follows: P=σ(Z) in S8-4, input node label matrix using one-hot encoding Each element of Y jc Represents node v j Whether it belongs to category c; construct cross entropy loss function Right now: S8-5. To prevent overfitting, add regularization loss Right now: where λ is used to control the regularization strength, ||·|| F is the Frobenius norm of the matrix; S8-6. Constructing the total loss function Update the classifier parameters: S8-7. Calculate the category probabilities of all unlabeled nodes and select the category with the highest probability as the classification result.

7. The method according to claim 1, characterized in that The step S9 includes the following sub-steps: S9-1. In the personalized matching system, there are M subject entities (such as users, devices, etc.) and N resource nodes, and a binary interaction matrix R∈{0,1} is constructed. M×N , R ui =1 represents the subject u to node v i There is interaction (such as click, collection, access, control, etc.), otherwise R ui =0; S9-2. For each subject entity u, average the embedding representations of all its interacted nodes to obtain the user vector e u , which is expressed as follows: in is the set of interactions of subject u, is node v i The node embedding of comes from the embedding matrix H in step S7; S9-3, in order to estimate the subject u's response to the node v i The preference degree of e u With h i Concatenate and input into a multi-layer perceptron for scoring: Among them [e u ;h i ] represents the concatenation operation of two vectors. MLP is a fully connected neural network with a nonlinear activation function. The output Represents the subject u to node v i preference score; S9-4. Use the stationary ranking loss function to train the recommendation model. The loss function is expressed as follows: It is a training set consisting of all known interaction samples; S9-5. For each subject entity u, calculate its prediction score for all candidate nodes Sort by scores from high to low, and select the top several nodes as the final personalized matching results.

Images