Multi-view primitive path convolutional network method based on residual attention

Abstract

The invention relates to the technical field of graph convolution, the core is a multi-view primitive path convolution network method and system, and the multi-view semi-supervised task performance is improved. The method comprises the following steps: 1, aiming at multi-view feature isomerism, calculating a similarity matrix of each view, and adding self-connection to nodes to obtain a symmetric normalized matrix; 2, combining view data integrity and task association degree, and calculating weight according to actual contribution; 3, in order to solve the problems of direct splicing dimension explosion and redundancy, self-adaptive splicing and efficient dimension reduction are carried out, and a low-dimension key fusion feature matrix is obtained; 4, aiming at the concealment of the key association in the traditional method, adaptively exploring a multi-length element path and performing weighted aggregation to obtain an accurate structural feature; and 5, a residual cross attention module is introduced during image convolution, and the problem of discrimination reduction caused by GCN stacking over-smoothing is solved.

Description

Technical Field

[0001] This invention relates to the fields of multi-view learning, graph convolutional networks, and attention mechanisms. Specifically, it is a graph convolutional network optimization method and system that introduces residual cross-attention modules and meta-paths into a multi-view learning framework. Background Technology

[0002] With the rapid development of multimedia technology and graph machine learning, multi-view learning has gradually become a core direction for processing complex high-dimensional data. Multi-view data originates from multiple information sources or possesses different feature forms, enabling a comprehensive characterization of the attributes and relationships of target objects from multiple perspectives. This overcomes the limitations of single-view data description and demonstrates key value in downstream tasks such as semi-supervised classification, clustering, and recommendation systems, powerfully promoting the practical application of multi-view-based learning methods in real-world scenarios.

[0003] Graph Convolutional Network (GCN), a classic technique in the field of graph learning, has become a research hotspot in graph structure data processing due to its powerful ability to aggregate topological neighbor features and extract high-order semantic features of nodes through multi-layer convolution operations. However, existing multi-view learning methods based on GCN generally have significant limitations: First, most methods assume that the graph constructed from multi-view data is isomorphic, assuming that all views reflect the same underlying entity relationships, but they ignore the heterogeneous nature of multi-view data—the graph structures corresponding to different views often represent different types of associations, resulting in the inability to effectively mine heterogeneous information; Second, existing heterogeneous graph learning methods rely heavily on predefined meta-paths to extract structural information, are overly dependent on domain knowledge, are difficult to adapt to different multi-view scenarios, and cannot dynamically adjust view weights to reduce interference from low-quality views; Third, multi-view data suffers from inconsistent feature dimensions, making it difficult for traditional methods to achieve efficient fusion of cross-view features. At the same time, traditional GCNs tend to become oversmooth as the number of network layers increases, making it difficult to stably explore higher-order cross-view relationships, and they lack an adaptive trade-off mechanism for meta-path information of different lengths. Therefore, the utilization of local structural information and global semantic information is insufficient, which seriously restricts the improvement of accuracy in multi-view tasks.

[0004] Therefore, how to break through the constraints of the isomorphic graph assumption, adaptively mine heterogeneous associations in multi-view data, efficiently solve the cross-view feature fusion problem, and achieve dynamic trade-offs between information at different levels has become a core issue that urgently needs to be addressed in multi-view GCN research. Based on this, this paper proposes a multi-view meta-path convolutional network method based on residual attention. By exploring multi-length cross-view meta-paths, introducing a meta-path graph-level attention mechanism, and combining a cross-attention module with residual connections, this method solves the pain points of existing multi-view GCNs in heterogeneity mining, feature fusion, and information trade-offs. This method can fully preserve the heterogeneous information of multi-view data while mitigating the negative impact of low-quality views and outliers on model performance through a dynamic optimization mechanism. Summary of the Invention

[0005] This invention proposes a multi-view graph convolutional network method and system based on residual cross-attention, the overall architecture of which is as follows: Figure 1 As shown, the flowchart is as follows Figure 2 This invention addresses the pain points of existing multi-view methods in terms of heterogeneity mining, feature fusion, and information trade-offs. It constructs an implicit meta-path graph matrix to automatically learn view weights, replacing predefined meta-paths and reducing reliance on domain knowledge. Simultaneously, it explores cross-view meta-paths of varying lengths to alleviate the oversmoothing limitations of traditional GCNs. Finally, it introduces a meta-path graph-level attention mechanism and a cross-attention module based on residual connections to adaptively allocate information weights for meta-paths of different lengths and mine deeper information, significantly improving the accuracy and stability of multi-view semi-supervised tasks.

[0006] The technology employed in this invention includes the following steps:

[0007] S1: Construct a relation matrix for each view and symmetrically normalize it. Considering the heterogeneous features of multi-view data, the k-nearest neighbor algorithm is used to obtain a similarity matrix for each view. Then, self-connections are added to each node, and the degree matrix of the nodes is calculated. Finally, the similarity matrix is ​​symmetrically normalized to obtain the adjacency matrix.

[0008] S2: Adaptive weighting of each view. To address the problem that traditional multi-view weighting often uses fixed values ​​or a single standard to determine weights, ignoring the differences in data quality among different views and resulting in high-contribution views not being utilized, an adaptive weighting method for multi-views is designed. First, the data integrity of each view and its relevance to the task objective are analyzed; then, the weight is calculated based on the view's contribution to the task result.

[0009] S3: Concatenate the feature matrix and reduce dimensionality. First, align the preprocessed features of each view according to a unified dimension and then concatenate them in an orderly manner to form a complete multi-view joint feature matrix. Then, a dimensionality reduction strategy is adopted to remove invalid and redundant information while retaining the core discriminative features, finally obtaining a low-dimensional fusion feature matrix rich in key information.

[0010] S4: Exploration and aggregation of multi-length meta-paths. First, based on the characteristics of the data structure of each view, multiple sets of meta-paths of different lengths are generated to mine multi-level correlation features. Then, weights are assigned by evaluating the actual contribution of each length meta-path to the task, and the correlation information of meta-paths of different lengths is weighted and fused to finally obtain comprehensive and accurate structural features.

[0011] S5: First, calculate the node embedding representation of the current layer based on the updated graph structure of the upper layer. Then, associate the node embedding representation with the multi-view features after initial concatenation and dimensionality reduction. The module then dynamically learns the weights of each view. The graph structure is then re-fused based on these weights as the graph structure input for the next convolutional layer. At the same time, the projection features are generated into residuals specific to this layer. Finally, the residual connections are used to supplement the information loss in the graph convolution, achieving consistent fusion of multi-view features.

[0012] Furthermore, step S1 specifically includes the following steps:

[0013] S11: We first use the k-nearest neighbor algorithm to determine the connection relationship between samples in the view, and obtain the similarity matrix of each view.

[0014] S12: Then, the adjacency matrix is ​​symmetrically normalized. This normalization step is to eliminate the influence of node degree differences on graph convolution operations, so that nodes of different degrees have a fair scale when feature aggregation, while allowing each node to contain its own information.

[0015] Furthermore, step S2 specifically includes the following steps:

[0016] S21: Given that different views contribute differently to downstream tasks, we adaptively learn the weight of each view in this step.

[0017] Furthermore, step S3 specifically includes the following steps:

[0018] S31: Concatenate all view feature matrices column-wise to obtain a high-dimensional feature matrix.

[0019] S32: High-dimensional feature concatenation is redundant, and excessively high dimensionality increases computational cost. Therefore, a multilayer perceptron is used to reduce the dimensionality of the high-dimensional feature matrix, compressing its dimension to a preset target dimension, to obtain a global feature matrix for extracting more compact global information.

[0020] Furthermore, step S4 specifically includes the following steps:

[0021] S41: To address the issues of traditional GCN's difficulty in handling heterogeneity and oversmoothing in multi-view data, as well as the difficulty in balancing local and global information, we obtain metapath matrices of different lengths by performing multiple matrix multiplications on the metapath matrix, which is essentially passing the relationships.

[0022] S42: After the previous step, we obtain different meta-path matrices. To address the lack of discriminative power caused by the traditional method treating all meta-path information equally, we then use a meta-path graph-level attention mechanism to adaptively assign weights to meta-path embeddings of different lengths for different samples, and finally fuse local and global information to obtain the final representation of the node. To ensure that the sum of the weights of all meta-paths of all lengths is 1, we normalize all attention weights.

[0023] S43: In order to assign weights separately to the embedding of each sample length to avoid interference between sample weights, we calculated the weight diagonal matrix.

[0024] S44: To obtain the final representation, we multiply the weight diagonal matrix of each length with the corresponding embedding to obtain the fused adjacency matrix.

[0025] Furthermore, step S5 specifically includes the following steps:

[0026] S51: The overall architecture of this part is as follows Figure 3 As shown, for each layer, we calculate the node embedding representation of the current layer by analyzing and fusing the graph structure of the previous layer and the initial fusion features, and then activate it through an activation function.

[0027] S52: To fuse the initial projection features of multiple views with the convolutional output of the previous layer, we need to obtain the importance weights of each view. This prepares for generating residuals and updating the graph structure, ultimately preventing node representation convergence in deep GCNs. We calculate the weights of each view in each layer, activate them using an activation function, and then normalize them.

[0028] S53: In order to achieve dynamic fusion where important views contribute more and secondary views update less, we fuse the features of all views by weighted summation to generate the fused features of the corresponding layer.

[0029] S54: In order to realize the residual structure update mechanism and control the fusion ratio of historical structure information and current multi-view structure information, first update the corresponding layer residual path matrix, then update and calculate the corresponding layer residual adjacency matrix, and finally calculate the latest graph structure matrix of the layer.

[0030] S55: In the output layer of the model, we obtain the final node features, which are the feature representations of the last layer.

[0031] S56: Finally, for this model, we define a formula for calculating the cross-entropy loss function applicable to multi-view semi-supervised tasks. This function serves as the optimization objective of the model. By minimizing this loss function, we adjust the model parameters to make the predicted probabilities as close as possible to the true labels. Attached Figure Description

[0032] The accompanying drawings of this invention are described below:

[0033] Figure 1 This is a model diagram of the multi-view meta-path convolutional network method based on residual cross-attention of the present invention;

[0034] Figure 2 This is a flowchart of the multi-view meta-path convolutional network method based on residual cross-attention of the present invention;

[0035] Figure 3 This is a diagram of the residual convolutional layer model of the present invention; Detailed Implementation

[0036] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0037] Please see Figures 1-3 ,like Figure 1 As shown, a preferred embodiment of the present invention provides a multi-view meta-path convolutional network method based on residual cross-attention, which mainly includes the following steps:

[0038] S1: Symmetric Normalization of Similarity Matrix: For multi-view data with heterogeneous features, the k-nearest neighbor algorithm is used to obtain the similarity matrix for each view. Then, self-connections are added to each node, and the degree matrix of the nodes is calculated. Finally, the similarity matrix is ​​symmetrically normalized to obtain the adjacency matrix. This normalization step aims to eliminate the influence of node degree differences on graph convolution operations, ensuring a fair scale for nodes with different degrees during feature aggregation.

[0039] S2: Adaptive Weighting of Each View: Addressing the common problem in traditional multi-view weighting methods that use fixed values ​​or a single standard to determine weights, ignoring differences in data quality across views (e.g., some views contain excessive noise or redundant information leading to weight imbalance), and failing to dynamically adjust weights based on specific task requirements, resulting in high-contribution views not being utilized, this paper designs an adaptive weighting method for multi-views. This method adjusts weights based on the data integrity of each view, its actual contribution to the task results, and its relevance to the task objectives. This strengthens the information proportion of effective views, suppresses interference from low-quality views, and achieves precise matching of view weights with data characteristics and task requirements.

[0040] S3: Feature Matrix Concatenation and Dimensionality Reduction: To address the issues of dimensionality explosion and increased computational burden on the model caused by direct concatenation of multi-view features, as well as the inconsistency in feature scales across different views and the presence of redundant information after concatenation that interferes with effective feature mining, an adaptive concatenation and efficient dimensionality reduction method for multi-view features is designed. First, the preprocessed features of each view are aligned according to a unified dimension and then concatenated in an orderly manner to form a complete multi-view joint feature matrix. Then, a dimensionality reduction strategy adapted to multi-source heterogeneous data is adopted to remove invalid and redundant information while retaining core discriminative features, balancing the integrity of feature representation with computational efficiency, and finally obtaining a low-dimensional fusion feature matrix rich in key information.

[0041] S4: Exploration and Aggregation of Multi-Length Meta-Paths: Addressing the issues of traditional multi-view meta-paths which only use fixed-length nodes for node association mining, losing differentiated information between direct associations of short paths and indirect associations of long paths, and failing to distinguish the contributions of meta-paths of different lengths during aggregation, leading to key associations being obscured by redundant information through simple merging, an adaptive exploration and weighted aggregation method for multi-length meta-paths is designed. First, based on the structural characteristics of each view's data, multiple sets of meta-paths of different lengths are generated to mine multi-level association features. Then, weights are assigned by evaluating the actual contribution of each length meta-path to the task, and the association information of meta-paths of different lengths is weighted and merged to ultimately obtain comprehensive and accurate structural features, providing richer association support for multi-view learning.

[0042] S5: Introducing Residual Cross-Attention: Graph convolution is performed, and a residual cross-attention module is introduced. The architecture of this part is as follows: Figure 3 As shown. In existing GCN-based multi-view learning, increasing the number of stacked layers can easily lead to oversmoothing, resulting in a decrease in the discriminative power of node representations. Furthermore, the lack of data-driven adaptation between multi-view features and graph fusion increases the training burden. Therefore, a residual cross-attention module is introduced.

[0043] First, the node embedding representation of the current layer is calculated based on the updated graph structure of the upper layer. Then, the node embedding representation is associated with the multi-view features after initial concatenation and dimensionality reduction, and the module dynamically learns the importance of each view. Next, the graph structure is re-fused according to this weight as the graph structure input of the next convolutional layer, and the projection features are generated into the residuals specific to this layer. Finally, the residual connections are used to supplement the information loss in the graph convolution, alleviate the oversmoothing phenomenon, achieve consistent fusion of multi-view features, and improve the performance of multi-view semi-supervised tasks.

[0044] Furthermore, step S1 specifically includes the following steps:

[0045] S11: We first use the k-nearest neighbor algorithm to determine the connectivity between samples in the view. We define the similarity matrix S of the v-th view. (v) as follows:

[0046]

[0047] in It is X (v) The i-th row vector, yes The set of k nearest neighbors.

[0048] S12: Then, the similarity matrix is ​​symmetrically normalized to obtain the adjacency matrix. First, to ensure that each node contains its own information, the similarity matrix S... (v) Adding the identity matrix I introduces a "self-loop". Normalization is used to eliminate the impact of node degree differences on graph convolution operations, ensuring a fair scale for nodes of different degrees during feature aggregation. Therefore, the adjacency matrix A of the v-th view... (v) The calculation formula is as follows:

[0049]

[0050] Where D (v) Represents the matrix similarity matrix S (v) The degree matrix corresponding to the matrix obtained by adding the identity matrix I.

[0051] S21: Because different views contribute differently to downstream tasks—for example, a view containing a lot of noise might offer little help for classification—we need to adaptively learn the weights of each view. In multi-view learning, adaptively assigning weights to each view means not pre-setting fixed values, but automatically learning and adjusting weight allocation based on dynamic characteristics such as the view's information quality, task suitability, and data integrity. The core objective is to give higher weights to high-value views, such as those with clear features, low noise, and strong information complementarity, and lower weights to low-quality views, such as those with occlusion, high noise, or redundancy, thereby optimizing the multi-view fusion effect. The fused view weights yield the meta-path matrix. The formula is as follows:

[0052]

[0053] Where v is the total number of views, β (v) For the weights of the v-th view, A (v) Let v be the normalized adjacency matrix of the v-th view.

[0054] S31: The core of feature matrix concatenation is to aggregate multi-dimensional, multi-modal, or multi-view information, allowing the model to simultaneously utilize feature patterns from different sources and improve its expressive power. Concatenating all view feature matrices column-wise yields a high-dimensional feature matrix X. The formula for calculating X is as follows:

[0055]

[0056] in d v Let V be the dimensional feature of the v-th view, where V is the total number of views and n is the number of samples.

[0057] S32: High-dimensional feature concatenation is redundant, and excessively high dimensionality increases computational costs. This invention uses a multilayer perceptron to reduce the dimensionality of X, compressing it to a preset target dimension. Obtain the global feature matrix in The calculation formula is as follows:

[0058]

[0059] This operation not only solves the problem of inconsistent dimensions, but also learns the non-linear relationships of multiple views and extracts more compact global information.

[0060] S41: To address the challenges of traditional GCNs in handling heterogeneity and oversmoothing in multi-view data, as well as the difficulty in balancing local and global information, we need to construct meta-paths. This is achieved by analyzing the meta-path matrix... Multiplying by itself q times, which corresponds to q steps of relational propagation, yields the q-length metapath matrix H.q H q The calculation formula is as follows:

[0061]

[0062] Where k is the maximum length of the metapath. For example, when q = 2, in The strength of the connection between sample i and sample j via a two-step meta-path represents the intensity of the connection.

[0063] S42: After step S41, we obtain k+1 metapath matrices {H}. 0 H 1 ,...,H k To address the insufficient discriminative power of traditional methods that treat all meta-path information equally, this invention employs a meta-path graph-level attention mechanism. This mechanism adaptively assigns weights to meta-path embeddings of different lengths for different samples, ultimately fusing local and global information to obtain the final representation of the node. The attention weight of the i-th node is calculated. The formula is as follows:

[0064]

[0065] in It is a feature under a q-length metapath, W q yes The weight matrix, b q Let W represent the bias vector, tanh(·) denote the activation function, and s be the global attention vector. q With b q It adapts to embedding feature distributions of different lengths, and uses s to measure the matching degree of the optimal feature of the embedding. The larger the value, the higher the matching degree, and the more important the metapath of that length is for the classification of sample i. Conversely, the smaller the value, the lower the matching degree and the less important it is.

[0066] To ensure that the weights of all length metapaths sum to 1, we calculate the sum of the attention scores for all metapaths. The softmax normalization formula is as follows:

[0067]

[0068] in for The corresponding value after normalization.

[0069] S43: To assign weights individually to the q-length embedding of each sample to avoid interference from sample weights, we need to obtain... Then construct the weight diagonal matrix Λ qThe off-diagonal elements of this matrix are 0, Λ q The formula is as follows:

[0070]

[0071] S44: To obtain the final representation Z, we will use the weight diagonal matrix Λ for each length. q With the corresponding embedded H q Multiplying them yields the merged adjacency matrix P, calculated using the following formula:

[0072]

[0073] S51: For the l-th layer, we first compute the node embedding representation Z of the current layer. (l) The calculation formula is as follows:

[0074]

[0075] in Represents the graph structure of the current layer, α l and β l W represents the trainable hyperparameters, and σ(·) represents the activation function. (l) For the trainable parameters of the graph convolutional layer, This represents the initial fusion feature of the l-th layer; the specific calculation formula will be introduced later.

[0076] S52: To fuse the initial projection features of multiple views with the convolutional output of the previous layer, we need to obtain the importance weights of each view. This prepares for generating specific residuals and updating the graph structure in subsequent layers, ultimately preventing node representation convergence in deep GCNs. For the view importance m of the next layer, i.e., the (l+1)th layer... (l+1) The calculation formula is as follows:

[0077]

[0078] Where σ(·) represents the activation function, This represents the number of nodes in the dataset, where i represents the nth node. This represents the vector of the i-th node output by the convolution of the l-th layer graph. This represents the multi-view feature matrix after projection of i nodes. q represents the learnable weight matrix corresponding to the initial projected features. (l+1) This represents the attention weight vector of the (l+1)th layer.

[0079] Then we need to deal with m. (l+1) After normalization, the weight percentage γ of each view is obtained. (l+1) The calculation formula is as follows:

[0080] γ(l+1) =softmax(m (l+1) ),

[0081] S53: In order to achieve dynamic fusion where important views contribute more and secondary views update less, we generate the fusion feature of the (l+1)th layer by weighted summation of the features of all views.

[0082]

[0083] in Represents the initial fusion feature of layer (l+1), and represents the multi-view aggregation result at the node attribute level of the current layer. This represents the features of the v-th view after dimensionality reduction. It allows the model to adaptively focus on views more valuable to the task, improving feature representation capabilities in multi-view scenarios.

[0084] S54: To implement the residual-based structural update mechanism, the path matrix of the residual element at layer l+1 is updated first. The calculation formula is defined as follows:

[0085]

[0086] Then update and calculate the residual adjacency matrix P of the (l+1)th layer. (l+1) The calculation formula is as follows:

[0087]

[0088] Finally, calculate the latest graph structure matrix at layer l. The calculation formula is as follows:

[0089]

[0090] Where α (l+1) This represents the learnable weights of the (l+1)th layer, controlling the fusion ratio of historical structural information and current multi-view structural information, and its value range is usually [0,1]. The core of this formula is to fuse the historical structure with the current structure of the multi-view.

[0091] S55: In the output layer of the model, we can obtain the final node feature Z', calculated using the following formula:

[0092] Z'=f(Z (L) )

[0093] Z (L) The final layer, or Lth layer, represents the feature representation, which consists of the node features after all graph convolutions, residuals, and other operations. f(·) is typically the softmax function, used to map the features to a class probability distribution.

[0094] S56: For this model, we define a cross-entropy loss function suitable for multi-view semi-supervised tasks. The calculation formula is as follows:

[0095]

[0096] Where Ω is the sample set, c is the total number of categories, and Y ij For real labels, Z' ij To be with Y ij The corresponding predicted probability is used as the optimization objective of the model. By minimizing this loss function, we adjust the model parameters to make the predicted probability as close as possible to the true label.

Claims

1. A multi-view meta-path convolutional network method based on residual attention, characterized in that, The method includes the following steps, implemented by a multi-view meta-path convolutional network method and system based on residual cross-attention. The overall architecture of the network and method is shown in Figures 1 and 2. The multi-view meta-path convolutional network method and system based on residual cross-attention includes constructing a relation matrix and corresponding weights, concatenating feature matrices and reducing dimensionality, exploring multi-length cross-view meta-paths, introducing a meta-path graph-level attention mechanism, and combining a cross-attention module with residual connections. The method includes: S1: Construct a relation matrix for each view and symmetrically normalize it. Considering the heterogeneous features of multi-view data, the k-nearest neighbor algorithm is used to obtain a similarity matrix for each view. Then, self-connections are added to each node, and the degree matrix of the nodes is calculated. Finally, the similarity matrix is ​​symmetrically normalized to obtain the adjacency matrix. S2: Adaptive weighting of each view. To address the problem that traditional multi-view weighting often uses fixed values ​​or a single standard to determine weights, ignoring the differences in data quality among different views and resulting in high-contribution views not being utilized, an adaptive weighting method for multi-views is designed. This method first analyzes the data integrity and relevance to the task objective of each view, and then calculates the weight based on the view's contribution to the task result. S3: Concatenate the feature matrix and reduce dimensionality. First, align the preprocessed features of each view according to a unified dimension and then concatenate them in an orderly manner to form a complete multi-view joint feature matrix. Then, a dimensionality reduction strategy is adopted to remove invalid and redundant information while retaining the core discriminative features, finally obtaining a low-dimensional fusion feature matrix rich in key information. S4: Exploration and aggregation of multi-length meta-paths. First, based on the characteristics of the data structure of each view, multiple sets of meta-paths of different lengths are generated to mine multi-level correlation features. Then, weights are assigned by evaluating the actual contribution of each length meta-path to the task, and the correlation information of meta-paths of different lengths is weighted and fused to finally obtain comprehensive and accurate structural features. S5: First, calculate the node embedding representation of the current layer based on the updated graph structure of the upper layer. Then, associate the node embedding representation with the multi-view features after initial concatenation and dimensionality reduction. The module then dynamically learns the weights of each view. The graph structure is then re-fused based on these weights as the graph structure input for the next convolutional layer. At the same time, the projection features are generated into residuals specific to this layer. Finally, the residual connections are used to supplement the information loss in the graph convolution, achieving consistent fusion of multi-view features.

2. The multi-view meta-path convolutional network method based on residual attention according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11: We first use the k-nearest neighbor algorithm to determine the connection relationship between samples in the view, and obtain the similarity matrix of each view. S12: Then, the adjacency matrix is ​​symmetrically normalized. This normalization step is to eliminate the influence of node degree differences on graph convolution operations, so that nodes of different degrees have a fair scale when feature aggregation, while allowing each node to contain its own information.

3. The multi-view meta-path convolutional network method based on residual attention according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21: Given that different views contribute differently to downstream tasks, we adaptively learn the weight of each view in this step.

4. The multi-view meta-path convolutional network method based on residual attention according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31: Concatenate all view feature matrices column-wise to obtain a high-dimensional feature matrix. S32: High-dimensional feature concatenation is redundant, and excessively high dimensionality increases computational cost. Therefore, a multilayer perceptron is used to reduce the dimensionality of the high-dimensional feature matrix, compressing its dimension to a preset target dimension, to obtain a global feature matrix for extracting more compact global information.

5. The multi-view meta-path convolutional network method based on residual attention according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41: To address the issues of traditional GCN's difficulty in handling heterogeneity and oversmoothing in multi-view data, as well as the difficulty in balancing local and global information, we obtain metapath matrices of different lengths by performing multiple matrix multiplications on the metapath matrix, which is essentially passing the relationships. S42: To address the lack of discriminative power caused by the traditional method treating all meta-path information equally, we use a meta-path graph-level attention mechanism to adaptively assign weights to different length meta-path embeddings for different samples, and finally fuse local and global information to obtain the final representation of the node. To ensure that the sum of the weights of all length meta-paths is 1, we normalize all attention weights. S43: In order to assign weights separately to the embedding of each sample length to avoid interference between sample weights, we calculated the weight diagonal matrix. S44: To obtain the final representation, we multiply the weight diagonal matrix of each length with the corresponding embedding to obtain the fused adjacency matrix.

6. The multi-view meta-path convolutional network method based on residual attention according to claim 1, characterized in that, Step S5 specifically includes the following steps: S51: The overall architecture of this part is shown in Figure 3. For each layer, we calculate the node embedding representation of the current layer by analyzing and fusing the graph structure of the previous layer and the initial fusion features, and then activate it through an activation function. S52: To fuse the initial projection features of multiple views with the convolutional output of the previous layer, we need to obtain the importance weights of each view. This prepares for generating residuals and updating the graph structure, ultimately preventing node representation convergence in deep GCNs. We calculate the weights of each view in each layer, activate them using an activation function, and then normalize them. S53: In order to achieve dynamic fusion where important views contribute more and secondary views update less, we fuse the features of all views by weighted summation to generate the fused features of the corresponding layer. S54: In order to realize the residual structure update mechanism and control the fusion ratio of historical structure information and current multi-view structure information, first update the corresponding layer residual path matrix, then update and calculate the corresponding layer residual adjacency matrix, and finally calculate the latest graph structure matrix of the layer. S55: In the output layer of the model, we obtain the final node features, which are the feature representations of the last layer. S56: Finally, for this model, we define a formula for calculating the cross-entropy loss function applicable to multi-view semi-supervised tasks. This function serves as the optimization objective of the model. By minimizing this loss function, we adjust the model parameters to make the predicted probabilities as close as possible to the true labels.

Images