A method for learning classification based on particle ball under multi-view data and related device

Abstract

The application discloses a granular ball-based graph learning classification method under multi-view data and a related device. The application obtains a multi-view data node set and generates a multi-view shared representation; based on the multi-view shared representation, the multi-view data node set is divided into a plurality of granular balls; based on the granular balls and an expert network, a routing preference vector of a node is obtained; based on the multi-view shared representation, a feature routing vector of the node is obtained; based on the routing preference vector and the feature routing vector, a routing weight of the node to each expert is obtained; through a plurality of expert networks, the multi-view shared representation is transformed to obtain a preliminary embedding representation of the node to each expert, and based on the routing weight of the node to each expert and the preliminary embedding representation of the expert, an embedding representation of the node is obtained; a prediction classification result of the multi-view data node is obtained according to the embedding representation of the node; and the application realizes accurate capture of fine-grained local structures in multi-source heterogeneous data, and significantly improves the representation ability and classification performance of the node.

Description

Technical Field

[0001] This application relates to the fields of graph neural networks and graph learning technology, specifically to a particle-based graph learning classification method and related apparatus for multi-view data. Background Technology

[0002] With the rapid development of multi-source information acquisition technologies, most research objects often exist in the form of multi-view data. For example, in social network analysis, the same user node can be described by multiple views, such as the friend relationship graph, the textual features of published content, and consumption behavior. Multi-view learning (MVL), by fusing complementary information from different sources, can more comprehensively and accurately improve the representation ability of objects than single-view models. Currently, graph convolutional networks (GCNs) have been widely used in the modeling and analysis of structured multi-view data, becoming a powerful tool for feature propagation through graph structures.

[0003] In practical engineering applications, although the complementary information provided by multiple views is beneficial to enhancing the robustness of the model, the high cost of annotation and the high requirements for domain expertise result in an extreme scarcity of labeled samples. Traditional supervised methods cannot utilize the vast majority of unlabeled samples, while classic semi-supervised methods struggle to effectively utilize the multi-view structure. Therefore, an efficient semi-supervised learning framework that can jointly utilize the complementarity of multiple views and unlabeled information has emerged. A series of studies have shifted their focus to the graph-based semi-supervised learning (GSSL) paradigm and extended this framework to multi-view scenarios. This involves constructing view-specific graphs, preserving the heterogeneous topology of each view, and seamlessly integrating nodes with topological information based on the smoothness assumption, enabling node propagation from labeled nodes to unlabeled nodes. However, extending graph semi-supervised learning to multi-view scenarios requires constructing dedicated graphs that preserve the heterogeneous topology of each view and designing view-specific or attention-driven aggregation mechanisms to fuse cross-view and local neighborhood information. The core challenge lies in balancing the effective fusion of multi-view representations with computational efficiency.

[0004] Multi-granularity computing provides a powerful framework for processing complex data at different levels. In the concept of granular-sphere computing, data points are encapsulated into spherical particles with a center and an adaptive radius, providing a coarse-grained representation that can capture local geometric structures while reducing computational complexity and improving the computational efficiency of clustering and classification tasks. Due to its structural modeling capabilities and computational efficiency, granular-sphere computing is naturally applied to graph-related tasks. This application introduces it into the multi-view graph construction process to adaptively capture local structural patterns at different scales while improving computational efficiency.

[0005] Currently, although existing technologies have achieved some success in applying GCN-based methods to multi-view semi-supervised scenarios, they still face several key challenges. First, existing technologies apply uniform transformations across all nodes and views, failing to effectively capture the inherent local structural patterns and diverse substructures in multi-view graphs. These fine-grained local patterns are crucial for effective representation learning. Second, existing technologies suffer from a trade-off between the effectiveness and efficiency of multi-view feature fusion computation. Simple feature concatenation strategies cannot capture cross-view complementarity, while complex fusion mechanisms incur excessively high computational costs. To address these issues, this application proposes a novel graph convolutional network architecture—a graph learning system and method for multi-view data based on granular-sphere computation and Mixture of Experts (MoE). Summary of the Invention

[0006] To address the issues of neglecting fine-grained local structural patterns and the contradiction between the effectiveness and efficiency of multi-view fusion in existing technologies, this application provides a graph learning classification method and related apparatus based on granular spheres for multi-view data.

[0007] To achieve the above objectives, this application provides the following technical solution:

[0008] The first aspect of this application provides a graph learning classification method based on spheres for multi-view data, including:

[0009] Retrieve multi-view data node set;

[0010] The multi-view data node set is input into a multi-view graph learning model to obtain the predicted classification results of the multi-view data nodes; the processing method in the multi-view graph learning model includes:

[0011] Generate a multi-view shared representation based on the multi-view data node set;

[0012] Based on the shared representation of multiple views, the set of multiple view data nodes is divided into several granular spheres;

[0013] Based on the partitioned spheres, the routing preference vectors of multi-view data nodes are obtained by combining them with expert networks; the feature routing vectors of multi-view data nodes are obtained based on the multi-view shared representation; and the routing weights of multi-view data nodes to each expert are obtained based on the routing preference vectors and feature routing vectors of multi-view data nodes.

[0014] The multi-view shared representation is transformed by an expert network to obtain the preliminary expert embedding representation of the multi-view data node. Based on the routing weights of each expert and the preliminary expert embedding representation of the multi-view data node, the node embedding representation is obtained.

[0015] The predicted classification results of multi-view data nodes are obtained based on the node embedding representation.

[0016] Furthermore, the method of dividing the multi-view data node set into several granular spheres based on multi-view shared representation is as follows:

[0017] Based on the shared representation of multiple views, an adaptive granular clustering algorithm is used to heuristically partition the set of data nodes of multiple views, generating several granular clusters with different granularities.

[0018] Furthermore, the routing preference vectors of multi-view data nodes are obtained by combining the partitioned spheres with the expert network; based on the routing preference vectors and feature routing vectors of the multi-view data nodes, the routing weights of the multi-view data nodes for each expert are obtained, specifically as follows:

[0019] set up E An expert, based on the division of granules and E The mapping relationship between experts is used to determine the preference matrix from the particle to the expert network, and then the routing preference vector of the multi-view data node is obtained, as specifically expressed below:

[0020]

[0021] in, Indicates the first The granular index to which each multi-view data node belongs. Preference matrix The OK, K Indicates the number of pellets. That is to say, the first All multi-view data nodes within the sphere to which a multi-view data node belongs are related to the entire... E A routing preference vector for each expert;

[0022] A gated network is configured, and the feature routing vectors of multi-view data nodes are obtained using multi-view shared representations, as specifically expressed below:

[0023]

[0024] in, Represents multi-view data nodes The feature routing vector; Represents multi-view data nodes Shared feature vectors; This represents the learnable parameters in the gated network; This represents the weight matrix of the first layer in a gated network; This represents the weight matrix of the second layer in the gated network; This represents the first-level bias term in the gated network; This represents the second-layer bias term in the gated network; Indicates the activation function;

[0025] The routing preference vector and feature routing vector of the multi-view data node are weighted and fused to obtain the routing weight of the multi-view data node for each expert, as specifically expressed below:

[0026]

[0027]

[0028] in, Represents multi-view data nodes The log-probability of the route; Indicates the control factor; Represents multi-view data nodes The routing weights for each expert.

[0029] Furthermore, the process of transforming the multi-view shared representation through an expert network to obtain the preliminary expert embedding representation of the multi-view data nodes specifically involves:

[0030]

[0031] in, Indicates the first Layered experts For multi-view data nodes Preliminary embedding characterization; Indicates the first Layered experts The learnable parameter matrix; Indicates the first Layered multi-view data nodes Shared feature vectors.

[0032] Furthermore, the node embedding representation obtained based on the routing weights of each expert and the preliminary embedding representation of the experts from the multi-view data nodes is specifically as follows:

[0033] Based on the routing weights of each expert across multi-view data nodes, the preliminary expert embedding representations output by all experts are weighted and aggregated to obtain the first... Layered multi-view data nodes The expert network output vector is specifically expressed as follows:

[0034]

[0035] in, Indicates the first Layered multi-view data nodes The expert network output vector; Experts For multi-view data nodes Routing weights;

[0036] The node embedding representation is obtained by propagating the expert network output vector through a graph structure using standard graph convolution.

[0037] Furthermore, the loss function used during the training of the multi-view graph learning model is a joint loss function that combines classification loss and load balancing loss; the classification loss is obtained based on the predicted classification results and their true labels of the labeled multi-view data nodes.

[0038] Furthermore, the load balancing loss specifically includes:

[0039]

[0040]

[0041] in, Indicates load balancing losses; This represents the average routing distribution of all multi-view data nodes; Indicates the first i Routing distribution of multiple view data nodes; 1 E It is an e-dimensional vector of size 1; KL represents relative entropy; Indicates the number of data nodes in the multi-view.

[0042] A second aspect of this application provides a particle-sphere-based graph learning classification system for multi-view data, comprising:

[0043] The shared representation module is used to acquire multi-view data node sets and generate multi-view shared representations;

[0044] The sphere partitioning module is used to divide a multi-view data node set into several spheres based on the multi-view shared representation;

[0045] The routing weight representation module is used to obtain the routing preference vector of multi-view data nodes based on the partitioned spheres and combined with the expert network; obtain the feature routing vector of multi-view data nodes based on the multi-view shared representation; and obtain the routing weight of multi-view data nodes to each expert based on the routing preference vector and feature routing vector of multi-view data nodes.

[0046] The node embedding representation module is used to transform the multi-view shared representation through the expert network to obtain the expert preliminary embedding representation of the multi-view data node, and to obtain the node embedding representation based on the routing weight of each expert for the multi-view data node and the expert preliminary embedding representation.

[0047] The classification module is used to obtain the predicted classification results of multi-view data nodes based on the node embedding representation.

[0048] A third aspect of this application provides a computer device, including: a processor and a computer-readable storage medium;

[0049] A processor, adapted to execute computer programs;

[0050] A computer-readable storage medium storing a computer program that, when executed by the processor, implements the above-described graph learning classification method based on spheres for multi-view data.

[0051] A fourth aspect of this application provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed by the above-described graph learning classification method based on spheres for multi-view data.

[0052] Compared with the prior art, this application has the following beneficial technical effects:

[0053] This application collaboratively processes multi-view graph data through methods such as multi-view shared representation generation, granular sphere partitioning, expert networks, and routing weight allocation, achieving efficient representation and accurate classification of multi-view graph data. This application constructs a shared representation using a multi-view autoencoder, mapping the original features of each view to a unified low-dimensional space. It captures cross-view information by minimizing reconstruction errors, abandoning complex and costly fusion mechanisms. While mining complementary information from multiple views, it significantly reduces model training and inference costs. By partitioning the multi-view node set into granular spheres and combining it with an expert network architecture, each expert is specifically adapted to feature transformations in specific granular sphere regions, significantly improving the analytical accuracy for diverse substructures in multi-view graphs and overcoming the limitation of existing technologies in effectively representing fine-grained local features. Subsequently, expert routing weights are calculated by combining routing preference vectors and node feature routing vectors. After transformation and weighted integration by the expert network, node embedding representations are obtained, and classification prediction is completed through these node embedding representations. This method achieves refined data processing, balanced allocation of expert resources, and efficient feature transformation, comprehensively optimizing model performance and adapting to the learning and classification needs of multi-source heterogeneous multi-view graph data. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1Flowchart of the particle-sphere-based graph learning classification method for multi-view data provided in this application;

[0056] Figure 2 This is a time efficiency comparison analysis chart in the embodiments of this application;

[0057] Figure 3 This is a schematic diagram of the structure of a particle-based graph learning classification system for multi-view data in an embodiment of this application;

[0058] Figure 4 This is a diagram showing the internal structure of a computer device in an embodiment of this application. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0060] Existing technologies face several key challenges in multi-view graph semi-supervised learning scenarios. First, labeled samples are scarce; traditional supervised methods cannot utilize the large number of unlabeled samples, while classic semi-supervised methods struggle to effectively leverage the multi-view structure. Second, extending graph semi-supervised learning to multi-view scenarios makes it difficult to balance effective fusion of multi-view representations with computational efficiency. Furthermore, existing technologies apply uniform transformations across all nodes and views, failing to effectively capture the inherent fine-grained local structural patterns and diverse substructures in multi-view graphs, which are crucial for effective representation learning. Their multi-view feature fusion also involves trade-offs: simple concatenation fails to capture cross-view complementarity, while complex fusion mechanisms incur excessive computational costs.

[0061] Based on this, in some embodiments of this application, such as Figure 1 As shown, a graph learning classification method based on spheres is provided for multi-view data, including the following steps:

[0062] Retrieve multi-view data node set;

[0063] The multi-view data node set is input into a multi-view graph learning model to obtain the predicted classification results of the multi-view data nodes; the processing method in the multi-view graph learning model includes:

[0064] Step 1: Generate a multi-view shared representation based on the multi-view data node set.

[0065] Specifically, obtain the contents M A multi-view data node set for each view, each view Corresponding to a specific feature matrix .

[0066] A multi-view autoencoder network is used to process the multi-view data node set to obtain a multi-view shared representation; the multi-view autoencoder network consists of a shared encoder. and M View decoder The composition and specific processing steps are as follows:

[0067] First, a shared encoder is used to project the inputs of each view into a unified low-dimensional space to obtain a multi-view shared representation matrix; then, a shared encoder is used... M The view decoder decodes the multi-view shared representation matrix to obtain... M The reconstruction feature matrix of each view is calculated as follows:

[0068]

[0069]

[0070] in, This indicates that multiple views share a representation matrix. It is a dimension of The multi-view shared representation matrix, its first... row vector That is to represent the first Shared feature vectors of multiple view data nodes; It is the identity matrix representing the node indices; It is the first m Reconstruction feature matrix of each view, Indicates a shared encoder; Indicates the first m The decoder corresponding to each view; Represents learnable encoder parameters; Indicates the first m Learnable decoder parameters corresponding to each view.

[0071] The multi-view autoencoder network is jointly trained by minimizing the total error of all views during the reconstruction process, as specifically calculated below:

[0072]

[0073] in, This represents the view reconstruction loss, which is... The reconstruction loss of each view is obtained by summing them. Indicates the first The reconstruction loss of each view Indicates the number of views. Indicates the number of data nodes in the multi-view dataset. Indicates the first The first view The original feature vectors of multiple view data nodes Indicates the first The first view Reconstructed feature vectors of multiple view data nodes.

[0074] This step involves setting up a multi-view autoencoder network with a single encoder and multiple decoders to extract and fuse consistent features across views from the original multi-source heterogeneous data, constructing a shared representation that integrates complementary information from multiple views. The multi-view autoencoder network mandates that the single shared multi-view representation must be able to reproduce the key features of all views. During training, the encoder automatically filters out random noise and redundant information in specific views, extracting only the core substructures that are consistent across multiple views. This lays a high-quality feature foundation for accurate particle segmentation and precise routing by hybrid experts in subsequent steps.

[0075] Step 2: Based on the multi-view shared representation, divide the multi-view data node set into several granular spheres;

[0076] Specifically, after obtaining the shared representation of all views, the multi-view data node set is heuristically partitioned using the adaptive Granular Ball Computing (GBC) algorithm to generate K spheres of different granularities. Each sphere represents a locally compact structure. Compared to traditional coarse-grained clustering, the spheres defined in this application can adaptively enclose the data by iteratively adjusting their radius based on the data distribution characteristics of the multi-view shared representation, thereby capturing more precise local boundaries.

[0077] It should be noted that the set of multi-view data nodes, containing all multi-view data nodes, can be regarded as an initial large-scale single sphere. Since the dispersion of multi-view data nodes within the sphere is extremely high in the initial state, it cannot describe any distribution characteristics of the data, thus serving as the starting point for splitting. Subsequently, an iterative splitting phase is entered, where the compactness of the sphere is calculated to determine whether to perform a split. Specifically, when the splitting operation improves the weighted average compactness of the generated sub-spheres, the sub-sphere is accepted. This process continues iteratively until the adaptive sphere clustering algorithm converges, i.e., the splitting operation can no longer optimize the overall compactness. After the compactness-based splitting ends, the adaptive sphere clustering algorithm dynamically calculates an adaptive radius threshold based on the current distribution of the sphere group. For spheres with radii exceeding the threshold, further refinement operations are performed to prevent the structure of local regions from being too coarse, ensuring the ability to capture fine features. Finally, very small spheres (less than 3 multi-view data nodes) are identified and merged into neighboring sub-spheres to ensure statistical reliability and filter noise. The above adaptive sphere clustering process produces a set of spheres. Among them, the first Multi-view data nodes Belongs to a cluster assignment ,in, K It is determined automatically, indicating that there is. Individual sphere division results This indicates a multi-view data node. The index of the constituent sphere, if Then it means the first Multi-view data nodes It has been officially classified as a granular material. The subset of nodes represented. Ultimately, multi-view data nodes. Based on their shared feature vectors Its position in space is uniquely classified into a subset of multi-view data nodes represented by a particular sphere.

[0078] This step performs clustering directly on the shared representation of the multi-view data, which incorporates complementary information from all views, rather than clustering individual views separately or performing clustering during training. This approach ensures that the discovered local structural patterns fully and consistently reflect the underlying distribution logic of the multi-view data. By adaptively adjusting the size and boundaries of the spheres, the inherent heterogeneous substructures in the multi-view graph can be captured, providing a scientific physical basis for assigning different graph convolution experts in subsequent steps.

[0079] Step 3: Based on the partitioned spheres, obtain the routing preference vector of the multi-view data node in combination with the expert network; obtain the feature routing vector of the multi-view data node based on the multi-view shared representation; obtain the routing weight of the multi-view data node for each expert based on the routing preference vector and feature routing vector of the multi-view data node.

[0080] Specifically, based on the determination in step 2 K Each ball, combined with a pre-set... E The mapping relationship between experts is used to determine the preference matrix from the particle to the expert network. The elements in the matrix This represents the affinity between the sphere k and the expert e, i.e., the selection preference of the multi-view data nodes within the sphere for a specific expert.

[0081] For the cluster assignment obtained in step 2 ,in express The index of the granular ball to which each multi-view data node belongs, for each multi-view data node. Furthermore, the routing preference vectors of the multi-view data nodes are obtained, and the specific calculation is as follows:

[0082]

[0083] in, Indicates the first The granular index to which each multi-view data node belongs. Preference matrix The OK, K Indicates the number of pellets. That is to say, the first All multi-view data nodes within the sphere to which a multi-view data node belongs are related to the entire... E A routing preference vector for each expert, which provides the expert prior distribution based on the membership of the sphere set.

[0084] To capture subtle feature differences among individual multi-view data nodes, a dedicated gated network is introduced. The feature routing vectors of the multi-view data nodes are calculated using the multi-view shared representation obtained in step 1. The feature routing vectors of individual multi-view data nodes are obtained based on the multi-view shared representation. These feature routing vectors are feature-aware dynamic routes that represent the feature-specific preferences of the multi-view data nodes. The specific calculation of the feature routing vectors of the multi-view data nodes is as follows:

[0085]

[0086] in, Represents multi-view data nodes The feature routing vector, The gated network directly accesses the multi-view data nodes. Shared feature vectors The calculation shows that the data node represents the multi-view data node. Based on its own attributes Individualized preferences of each expert. This represents the learnable parameters in the gated network; This represents the weight matrix of the first layer in a gated network; This represents the weight matrix of the second layer in the gated network; These are the learnable parameters of a two-layer multilayer perceptron (MLP) in a gated network. This represents the first-level bias term in the gated network; This represents the second-layer bias term in the gated network; It is the bias vector in a neural network; The activation function is represented by ReLU (Rectified Linear Unit), a non-linear activation function, which is used to introduce non-linearity so that gating can capture complex feature preferences.

[0087] The route log odds are obtained by weighted fusion of the route preference vector and feature route vector of the multi-view data nodes, as calculated as follows:

[0088]

[0089] in, Represents multi-view data nodes The log-probability of the route; Indicates control factor. Used for regulation and The balance between them.

[0090] Subsequently, the log-probability of the merged routes is normalized using the softmax function, transforming it into a probability distribution form to obtain the routing weights of the multi-view data nodes for each expert. The specific calculation is as follows:

[0091]

[0092] in, Represents multi-view data nodes The routing weights for each expert, The specification defines the multi-view data nodes during the next layer of processing. The proportion of information flowing to each expert.

[0093] This step constructs a mapping and routing mechanism between the particle and the expert network by introducing... A weighted mechanism, comprehensively considering regional features and individual multi-view data node features, forms adaptive routes for different local patterns, realizing a fine-grained expert selection mechanism. In real-world multi-view scenarios, single dynamic feature gating is highly susceptible to topological conflicts between different views, leading to routing decision failure. This step introduces granular spherical regional features generated based on multi-view shared representations, providing a cross-view reflection for routing decisions. Figure 1 A reliable benchmark for consistent structure corrects routing deviations caused by noise in individual views. By synergistically fusing this structural benchmark with the characteristics of multi-view data nodes, it ensures that multi-view data nodes belonging to the same granularity can be accurately and stably assigned to the most relevant expert combinations for processing. This fine-grained expert selection mechanism drives each expert network to deeply mine specific substructure patterns in heterogeneous graph data, thereby fundamentally improving the model's parsing accuracy for complex multi-source data.

[0094] Step 4: Transform the multi-view shared representation through the expert network to obtain the expert preliminary embedding representation of the multi-view data node. Based on the routing weights of each expert and the expert preliminary embedding representation of the multi-view data node, obtain the node embedding representation. Obtain the prediction classification result of the multi-view data node based on the node embedding representation.

[0095] Specifically, in the expert pool, each expert Each maintains a set of independent learnable parameter matrices Regarding the first Layered multi-view data nodes Given the input features, each expert applies their specific linear transformation to obtain multi-view data nodes from the expert's perspective. The initial embedding representation by experts is calculated as follows:

[0096]

[0097] in, Indicates the first Layered experts For multi-view data nodes The initial embedding representation, representing multi-view data nodes In the Intermediate states generated under the processing of multiple experts; Indicates the first Layered experts The learnable parameter matrix represents the first... The first in the layer Each expert has learnable parameters that are responsible for performing linear transformations on the input features from a specific perspective. Indicates the first Layered multi-view data nodes The shared feature vector, when the initial layer ( When this value is obtained, it is the multi-view shared representation matrix extracted in step 1. Multi-view data nodes Shared feature vectors .

[0098] The routing weights of each expert are based on the multi-view data nodes generated in step 3. We perform weighted aggregation on the preliminary expert embedding representations output by all experts to obtain the first... Layered multi-view data nodes expert network output vector This leads to the output matrix obtained after expert network aggregation. ; The specific expression is as follows:

[0099]

[0100] in, Indicates the first Layered multi-view data nodes The expert network output vector represents the first... Layered multi-view data nodes The integrated features resulting from the fusion of multiple local patterns; Experts For multi-view data nodes The feature routing weights determine the multi-view data nodes. For the first The degree to which an expert's processing results are relied upon or adopted.

[0101] Aggregated expert output matrix With normalized adjacency matrix In combination, feature information is propagated in the graph structure through standard graph convolution operations, specifically as follows:

[0102]

[0103] in, normalized adjacency matrix The elements in the table represent multi-view data nodes. With neighboring nodes Topological connection weights between them; For multi-view data nodes ; Represents multi-view data nodes Nodes in the neighborhood; For the first Nodes in the layer neighborhood The output vector after aggregation by the expert network; For activation functions; No. Layered multi-view data nodes Shared feature vectors.

[0104] The first layer of the expert network uses the nonlinear activation function ReLU and dropout to generate hidden layer embeddings with local neighborhood information. To further extract higher-order features, a second layer of the expert network is constructed, which embeds nodes in the hidden layer... The process of repeated aggregation and graph propagation ultimately generates node embedding representations. , , They are as follows:

[0105]

[0106]

[0107] in, As the first layer output, it represents the embedding representation of the hidden layer. The hidden layer embedding incorporates local neighborhood information and incorporates the Dropout (0.5) mechanism, which randomly deactivates 50% of neurons during training to prevent the model from overfitting. The output of the second layer represents the node embedding representation. Integrated multi-view Figure 1 Consistency, local boundary characteristics of particles and spheres, and global topological relationships constitute the final semantic features in high-dimensional space; This represents the output matrix after all multi-view data nodes in the first layer have been aggregated by the expert network; This represents the output matrix after all multi-view data nodes in the second layer have been aggregated by the expert network.

[0108] Embedded features of nodes Input the classification layer to obtain the predicted labels for the multi-view data node set. The specific calculations are as follows:

[0109]

[0110] in, This represents the predicted label probability distribution of the multi-view data node set; LogSoftmax is used as the activation function to map the high-dimensional feature vector to the log probability of each category, and the model will finally select the category with the highest probability as the classification result.

[0111] This step assigns differentiated expert parameters to multi-view data nodes in different sphere regions, enabling the multi-view graph learning model to adaptively adapt to diverse substructure patterns in heterogeneous multi-view graphs. Simultaneously, a hybrid expert mechanism is used to capture the heterogeneity of local features of multi-view data nodes, while GCN captures the consistency of topological structures. The deep integration of these two approaches significantly improves classification accuracy and robustness in semi-supervised scenarios where labels are extremely scarce.

[0112] In some embodiments of this application, the loss function used during the training of the multi-view graph learning model is a joint loss function combining classification loss and load balancing loss; the classification loss is obtained based on the predicted classification results and true labels of the labeled multi-view data nodes. Based on the classification loss, the joint loss of the multi-view data model is obtained by combining the load balancing loss; the parameters of the multi-view graph learning model are iteratively updated based on the joint loss; the nodes to be classified are embedded into the updated multi-view graph learning model using the representation and graph structure inputs to obtain the predicted classification results of the multi-view data nodes to be classified.

[0113] Specifically, we first construct the joint loss function. The joint loss of the multi-view graph learning model consists of the classification loss and the load balancing loss, as shown below:

[0114]

[0115] in, Indicates joint loss, This represents the balancing hyperparameter, used to adjust the weight ratio between the classification task and the load balancing constraint; Indicates classification loss; This indicates load balancing losses.

[0116] It is for a set of labeled multi-view data nodes. The negative log-likelihood loss between the predicted result and the true label is calculated as follows:

[0117]

[0118] in, This indicates that the set of multi-view data nodes has been marked. S Indicates the number of categories, It's a real label. Represents multi-view data nodes Category The predicted probability.

[0119] To ensure balanced utilization of the expert network and prevent only a few experts from dominating, a load balancing loss is introduced to prevent multi-view graph learning models from using only one or two experts while ignoring other experts. The specific calculations are as follows:

[0120]

[0121] in, This represents the average routing distribution of all multi-view data nodes. ;1 E It is an e-dimensional vector of 1, representing that each expert bears exactly... The ideal state of task quantity; KL stands for relative entropy (Kullback-Leibler Divergence), which is a metric used to measure the difference between two distributions. Here, it is used to measure the degree of deviation between the actual expert network assignment and the ideal uniform assignment.

[0122] After obtaining the joint loss, the gradient of the joint loss with respect to the parameters of each module in the multi-view graph learning model is calculated using the backpropagation algorithm. The parameters in the multi-view graph learning model are iteratively updated using optimizers such as Adam (Adaptive Moment Estimation, an optimization algorithm used to train deep learning models) or SGD (Stochastic Gradient Descent) until the multi-view graph learning model converges.

[0123] The node embedding representation and graph structure to be classified are input into the updated multi-view graph learning model to obtain the predicted classification result of the multi-view data nodes to be classified. Specifically, this is achieved through the node embedding features finally generated in step 4. The output layer calculates the probability distribution of multi-view data nodes belonging to each category, and selects the category with the highest probability as the final classification result of the unlabeled multi-view data nodes.

[0124] This step, through end-to-end joint training, enables the four stages of shared representation, particle partitioning, expert routing weights, and feature transformation to evolve collaboratively, ensuring that the geometric structure captured by the particles truly serves to improve the accuracy of the classification task. Furthermore, building upon step 4's use of the expert network to endow the model with the ability to handle heterogeneous substructures in different particle regions, a load balancing loss is introduced. The routing of experts is macroscopically controlled to avoid the risk of pattern collapse caused by the over-utilization of some experts, ensuring that each local pattern can obtain the most sufficient feature mining.

[0125] In summary, this application constructs a multi-view shared representation that integrates complementary information from multiple views, utilizes adaptive granular clustering to accurately capture fine-grained local structures in the feature space, and establishes an expert routing weight mechanism based on local geometric structures, thereby achieving adaptive modeling of region-aware expert networks. Finally, end-to-end training is performed using a joint loss function that includes load balancing constraints to achieve deep feature extraction and accurate classification of multi-view data.

[0126] Example

[0127] Dataset selection and preprocessing: To verify the wide applicability of this method in different disciplines, this embodiment selects 6 publicly available multi-view benchmark datasets. The statistical information is shown in Table 1. These datasets cover the two major domains of text and image. All datasets follow the semi-supervised learning setting: 10% of the samples are randomly selected as labeled data (class balanced partitioning), and the remaining 90% are unlabeled data. During the training process, only labeled samples are relied upon, and the classification performance is finally evaluated on unlabeled samples.

[0128] Table 1. Statistics of the Multi-View Benchmark Dataset

[0129]

[0130] Dataset preprocessing instructions:

[0131] Text datasets (BBCnews, BBCsports): After word segmentation and stop word removal of the original text, feature vectors are extracted using TF-IDF to ensure that the feature dimensions of each view are uniform to the same order of magnitude.

[0132] Image datasets (HW, MNIST, MNIST10k, MSRC-v1): Images are normalized (pixel values ​​are mapped to [0,1]), and multi-view features are generated using different feature extractors (such as HOG, LBP, and CNN shallow features).

[0133] All view features are standardized (mean 0, variance 1) to eliminate the impact of dimensional differences on model training.

[0134] In this embodiment, the core parameter configuration of the multi-view graph learning model is as follows: In the multi-view autoencoder network, the shared encoder is a 2-layer perceptron (MLP) (input dimension = sum of feature dimensions of each view, hidden dimension = 512, output dimension = 128), each view corresponds to a 1-layer MLP decoder (input dimension = 128, output dimension = original feature dimension of the corresponding view), the reconstruction loss uses the Adam optimizer (learning rate = 1e-3), and the training epochs = 200; the adaptive sphere clustering initially consists of a single sphere containing all multi-view data nodes, the splitting stopping condition is that the weighted average compactness improvement of the sub-spheres is ≤ 0.001, and the minimum number of multi-view data nodes in the sphere is set to 3 (if less than this number, they are merged into the nearest neighbor sphere); in the expert network, the number of experts E = 4, the gating network is a 2-layer MLP (input dimension = 128, hidden dimension = 64, output dimension = 4), and the clustering-feature fusion coefficient is... The datasets were adapted according to their characteristics (BBCnews=0.7, BBCsports=0.9, HW=0.5, MNIST=0.8, MNIST10k=0.9, MSRC-v1=0.3); the graph convolutional network was a 2-layer structure (the first layer included Dropout=0.5), the optimizer was Adam (weight decay=1e-4), the learning rate was 1e-3, the cosine annealing strategy was used during training, the number of training epochs T=300, and the load balancing loss weight λ=0.1 (λ=1.0 for the MNIST10k dataset); the experimental evaluation metrics were classification accuracy (ACC) and F1 score (harmonic mean of precision and recall), and all experiments were repeated 6 times (mean ± standard deviation) to ensure the reliability of the results.

[0135] Comparison Methods: Eight mainstream multi-view semi-supervised learning methods were selected as baselines, covering different types such as traditional graph methods, single-view GCN variants, and multi-view fusion GCNs to ensure comprehensiveness and representativeness of the comparison. Specifically, these include: LP-Multi (Label Propagation with Multi-feature Fusion); GCN-Multi (Multi-view Graph Convolutional Network), a graph learning method that trains each view independently using GCNs and then fuses the results; Co-GCN (Co-training Graph Convolutional Network); IMvGCN (Interactive Multi-view Graph Convolutional Network); DSRL (Discriminative Sparse Representation Learning); LEGCN-FF (Learnable Graph Convolution Network with Feature Fusion); and JEGCN (Joint Embedding and Topology-adaptive Fusion Graph Convolutional Network). Network (Joint Embedding and Topology Adaptive Fusion Graph Convolutional Network); GEGCN (Graph Ensemble Graph Convolutional Network).

[0136] Classification performance comparison: The comparison results of ACC and F1 scores of all the above methods on 6 datasets are shown in Table 2 below:

[0137] Table 2 Classification performance comparison data

[0138]

[0139] Experimental results demonstrate that this embodiment exhibits significant comprehensive advantages in multi-view semi-supervised node classification tasks: achieving the best ACC on BBCsports, HW, MNIST, and MNIST10k datasets; slightly higher than GEGCN on MSRC-v1; and lower than JEGCN and GEGCN on BBCnews, but with a standard deviation of 0, exhibiting stronger stability. Compared to traditional methods (such as GCN-Multi), this embodiment achieves significant performance improvements by accurately capturing local structures through granular-sphere clustering and adapting heterogeneous subgraphs using a hybrid expert architecture, with a 2.29% improvement in ACC on the HW dataset and a 0.98% improvement on MNIST10k. Compared to advanced multi-view GCN methods (JEGCN, GEGCN), this embodiment shows more pronounced advantages on large-scale datasets (MNIST10k) and complex image datasets (HW), fully validating its effectiveness in handling high-dimensional heterogeneous data. Notably, the standard deviation of this embodiment is close to 0 on all datasets, indicating that the routing mechanism guided by granular-sphere clustering effectively enhances the model's robustness and successfully alleviates the instability caused by multi-view topological conflicts.

[0140] Ablation Experiment Results and Analysis: To verify the necessity of the core components, this embodiment designed two model variants for ablation experiments: with / o MoE with the hybrid expert architecture removed and replaced by a single GCN layer, and with / o Cluster with the adaptive granular clustering removed and only the routing weights of the feature-gated experts retained.

[0141] The performance comparison results of the ablation experiments show that the full-configuration model in this embodiment achieves the best performance on both the HW and MSRC-v datasets: On the HW dataset, the ACC with / o MoE is 95.73±0.20%, and the ACC with / o Cluster is 95.03±0.47%, which are 0.17% and 0.87% lower than the full-configuration model, respectively; On the MSRC-v1 dataset, the ACC with / o MoE is 82.43±2.18%, and the ACC with / o Cluster is 81.90±3.35%, which are 2.97% and 3.5% lower than the full-configuration model, respectively.

[0142] Analysis shows that removing particle-sphere clustering (without clustering) generally degrades model performance, especially on MSRC-v1 where the ACC drops by 3.5%. This indicates that cluster-guided structural priors can effectively correct routing biases in feature gating and improve expert assignment accuracy. The with-MoE variant exhibits slight performance fluctuations on some datasets and significant performance degradation on others, suggesting that a single GCN can partially replace the hybrid expert architecture on simple datasets, but still has limitations on complex datasets. In summary, adaptive particle-sphere clustering and the hybrid expert architecture are the core guarantees for achieving high performance in this application.

[0143] Clustering method comparison experiment: To verify the superiority of GBC, it was compared with traditional clustering methods (K-means clustering algorithm, DBSCAN (Density-Based Spatial Clustering of Applications with Noise)). All three clustering methods are based on shared representations of the same dimension: K-means has a cluster number k=8, which is set as a medium number of clusters, providing a good balance between granularity and computational efficiency. The ε parameter of DBSCAN is automatically estimated by calculating the median distance from the sample to the 10th nearest neighbor, and the minimum number of samples is set to 3.

[0144] Performance comparison results show that GBC in this embodiment outperforms K-means and DBSCAN on all datasets: On the BBCnews and BBCsports datasets, where the data structures are relatively simple, all three methods achieved competitive performance; on the MNIST dataset, GBC's ACC is 90.68±0.00% and F1 is 88.98±0.00%, significantly better than K-means (84.18% ACC, 76.31% F1) and DBSCAN (83.52% ACC, 75.52% F1); on the MNIST10k dataset, GBC also maintained a performance advantage of about 5%.

[0145] Furthermore, visualization of clustering results using t-SNE (t-distributed stochastic neighbor embedding) reveals that GBC can generate fine-grained micro-clusters, effectively preserving local sample similarity and providing a tight community structure for routing. In contrast, K-means and DBSCAN, constrained by global optimization objectives, tend to merge heterogeneous samples into the same cluster, leading to expert assignment errors and impacting model performance. GBC's advantages are particularly pronounced on complex datasets, fully validating its unique ability to capture local geometric structures in high-dimensional data.

[0146] Computational efficiency analysis: such as Figure 2 As shown, our method demonstrates good time efficiency on all evaluated datasets. On smaller datasets such as HW and MSRC-v1, our method requires comparable or less training time than most multi-view graph baselines while achieving higher or competitive classification accuracy. On larger datasets, particularly MNIST10k, although some methods significantly increase computational cost, our method maintains a moderate training time while delivering strong performance. This demonstrates that our method offers a good balance between effectiveness and efficiency.

[0147] In one embodiment of this application, such as Figure 3 As shown, a graph learning classification system based on spheres for multi-view data is provided, including:

[0148] The shared representation module is used to acquire multi-view data node sets and generate multi-view shared representations;

[0149] The sphere partitioning module is used to divide a multi-view data node set into several spheres based on the multi-view shared representation;

[0150] The routing weight representation module is used to obtain the routing preference vector of multi-view data nodes based on the partitioned spheres and combined with the expert network; obtain the feature routing vector of multi-view data nodes based on the multi-view shared representation; and obtain the routing weight of multi-view data nodes to each expert based on the routing preference vector and feature routing vector of multi-view data nodes.

[0151] The node embedding representation module is used to transform the multi-view shared representation through the expert network to obtain the expert preliminary embedding representation of the multi-view data node, and to obtain the node embedding representation based on the routing weight of each expert for the multi-view data node and the expert preliminary embedding representation.

[0152] The classification module is used to obtain the predicted classification results of multi-view data nodes based on the node embedding representation.

[0153] For specific limitations regarding the particle-sphere-based graph learning classification system for multi-view data, please refer to the limitations of the particle-sphere-based graph learning classification method for multi-view data mentioned above. The corresponding technical effects can be obtained equally and will not be repeated here. Each module in the aforementioned particle-sphere-based graph learning classification system for multi-view data can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0154] Figure 4 An internal structural diagram of a computer device is shown in one embodiment. This computer device may specifically be a terminal or a server. Figure 4As shown, the computer device includes a processor, memory, network interface, display, camera, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a particle-based graph learning classification method for multi-view data. The display screen can be an LCD screen or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0155] Those skilled in the art will understand that Figure 4 The computer device structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computing devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0156] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described graph learning classification method based on particle spheres under multi-view data.

[0157] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described graph learning classification method based on spheres for multi-view data.

[0158] In summary, this application provides a graph learning classification method, system, computer device, and storage medium based on granular spheres for multi-view data. This application focuses on the task of node classification in multi-view data under graph semi-supervised learning, using granular sphere partitioning as an intermediate step for adaptively capturing local spatial units, and deeply integrating a hybrid expert mechanism with a graph convolutional network architecture. This design addresses the real-world scenario where multi-view labels are extremely scarce. By introducing graph neural networks and adjacency matrices to process graph data with complex structured relationships (such as node interactions in social network analysis or bioinformatics prediction), it overcomes the bottleneck of traditional methods that only process independent sample sets by utilizing feature propagation and semantic mining. While achieving accurate capture of fine-grained local structures, it ensures the professionalism and balance of multi-view feature extraction.

[0159] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0160] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this application, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A graph learning classification method based on spheres for multi-view data, characterized in that, include: Retrieve multi-view data node set; The multi-view data node set is input into the multi-view graph learning model to obtain the predicted classification results of the multi-view data nodes; The processing methods in the multi-view graph learning model include: Generate a multi-view shared representation based on the multi-view data node set; Based on the shared representation of multiple views, the set of multiple view data nodes is divided into several granular spheres; Based on the partitioned spheres, the routing preference vectors of multi-view data nodes are obtained in conjunction with the expert network; the feature routing vectors of multi-view data nodes are obtained based on the multi-view shared representation; and the routing weights of multi-view data nodes to each expert are obtained based on the routing preference vectors and feature routing vectors of multi-view data nodes, specifically: set up E An expert, based on the division of granules and E The mapping relationship between experts is used to determine the preference matrix from the particle to the expert network, and then the routing preference vector of the multi-view data node is obtained, as specifically expressed below: in, Indicates the first The granular index to which each multi-view data node belongs. Preference matrix The OK, K Indicates the number of pellets. That is to say, the first All multi-view data nodes within the sphere to which a multi-view data node belongs are related to the entire... E A routing preference vector for each expert; A gated network is configured, and the feature routing vectors of multi-view data nodes are obtained using multi-view shared representations, as specifically expressed below: in, Represents multi-view data nodes Feature routing vector; Represents multi-view data nodes Shared feature vectors; This represents the learnable parameters in the gated network; This represents the weight matrix of the first layer in a gated network; This represents the weight matrix of the second layer in the gated network; This represents the first-level bias term in the gated network; This represents the second-layer bias term in the gated network; Indicates the activation function; The routing preference vector and feature routing vector of the multi-view data node are weighted and fused to obtain the routing weight of the multi-view data node for each expert, as specifically expressed below: in, Represents multi-view data nodes The log-probability of the route; Indicates the control factor; Represents multi-view data nodes The routing weights for each expert; The multi-view shared representation is transformed by an expert network to obtain the preliminary expert embedding representation of the multi-view data node. Based on the routing weights of each expert and the preliminary expert embedding representation of the multi-view data node, the node embedding representation is obtained. The predicted classification results of multi-view data nodes are obtained based on the node embedding representation.

2. The graph learning classification method based on spheres for multi-view data according to claim 1, characterized in that, The method of dividing the multi-view data node set into several granular spheres based on multi-view shared representation is as follows: Based on the shared representation of multiple views, an adaptive granular clustering algorithm is used to heuristically partition the set of data nodes of multiple views, generating several granular clusters with different granularities.

3. The graph learning classification method based on spheres for multi-view data according to claim 1, characterized in that, The process of transforming the multi-view shared representation through an expert network to obtain the preliminary expert embedding representation of the multi-view data nodes specifically involves: in, Indicates the first Layered experts For multi-view data nodes Preliminary embedding characterization; Indicates the first Layered experts The learnable parameter matrix; Indicates the first Layered multi-view data nodes Shared feature vectors.

4. The graph learning classification method based on spheres for multi-view data according to claim 3, characterized in that, The node embedding representation obtained based on the routing weights of each expert and the preliminary embedding representation of the experts from the multi-view data nodes is as follows: Based on the routing weights of each expert across multi-view data nodes, the preliminary expert embedding representations output by all experts are weighted and aggregated to obtain the first... Layered multi-view data nodes The expert network output vector is specifically expressed as follows: in, Indicates the first Layered multi-view data nodes The expert network output vector; Experts For multi-view data nodes Routing weights; The node embedding representation is obtained by propagating the expert network output vector through a graph structure using standard graph convolution.

5. The graph learning classification method based on spheres for multi-view data according to claim 1, characterized in that, The loss function used during the training of the multi-view graph learning model is a joint loss function that combines classification loss and load balancing loss; the classification loss is obtained based on the predicted classification results and their true labels of the labeled multi-view data nodes.

6. The graph learning classification method based on spheres for multi-view data according to claim 5, characterized in that, The load balancing loss is specifically as follows: in, Indicates load balancing losses; This represents the average routing distribution of all multi-view data nodes; Indicates the first i Routing distribution of multiple view data nodes; 1 E It is an e-dimensional vector of size 1; KL represents relative entropy; Indicates the number of data nodes in the multi-view.

7. A graph learning classification system based on spheres for multi-view data, characterized in that, include: The shared representation module is used to acquire multi-view data node sets and generate multi-view shared representations; The sphere partitioning module is used to divide a multi-view data node set into several spheres based on the multi-view shared representation; The routing weight representation module is used to obtain the routing preference vector of multi-view data nodes based on the partitioned spheres and combined with the expert network; to obtain the feature routing vector of multi-view data nodes based on the multi-view shared representation; and to obtain the routing weight of multi-view data nodes for each expert based on the routing preference vector and feature routing vector of multi-view data nodes. Specifically: set up E An expert, based on the division of granules and E The mapping relationship between experts is used to determine the preference matrix from the particle to the expert network, and then the routing preference vector of the multi-view data node is obtained, as specifically expressed below: in, Indicates the first The granular index to which each multi-view data node belongs. Preference matrix The OK, K Indicates the number of pellets. That is to say, the first All multi-view data nodes within the sphere to which a multi-view data node belongs are related to the entire... E A routing preference vector for each expert; A gated network is configured, and the feature routing vectors of multi-view data nodes are obtained using multi-view shared representations, as specifically expressed below: in, Represents multi-view data nodes Feature routing vector; Represents multi-view data nodes Shared feature vectors; This represents the learnable parameters in the gated network; This represents the weight matrix of the first layer in a gated network; This represents the weight matrix of the second layer in the gated network; This represents the first-level bias term in the gated network; This represents the second-layer bias term in the gated network; Indicates the activation function; The routing preference vector and feature routing vector of the multi-view data node are weighted and fused to obtain the routing weight of the multi-view data node for each expert, as specifically expressed below: in, Represents multi-view data nodes The log-probability of the route; Indicates the control factor; Represents multi-view data nodes The routing weights for each expert; The node embedding representation module is used to transform the multi-view shared representation through the expert network to obtain the expert preliminary embedding representation of the multi-view data node, and to obtain the node embedding representation based on the routing weight of each expert for the multi-view data node and the expert preliminary embedding representation. The classification module is used to obtain the predicted classification results of multi-view data nodes based on the node embedding representation.

8. A computer device, characterized in that, include: Processor and computer-readable storage media; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program that, when executed by the processor, implements the particle-based graph learning classification method for multi-view data as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in any one of claims 1 to 6: a particle-based graph learning classification method for multi-view data.

Images