Conditional graph and dual-view decoupled graph structure representation learning method and system

Abstract

The application relates to the technical field of graph structure data processing, and discloses a graph structure representation learning method and system based on conditional graphs and double-view decoupling, which comprises the following steps: extracting node position representation of nodes in an original graph view, constructing a structure condition vector of each node according to node degrees, neighbor average degrees and graph densities based on statistics of the original graph view, constructing a structure condition matrix, and then constructing an auxiliary graph view; inputting the node position representation in the original graph view and the auxiliary graph view, an original graph normalized adjacency matrix and an auxiliary graph normalized adjacency matrix into a trained double-view representation learning model to perform double-view node coding, obtaining node local embedding, original graph view structure embedding and auxiliary graph view structure embedding, and finally obtaining final node representation after fusion; and performing a downstream task based on the final node representation. The information integrity and structure expression capability of the node representation can be improved, so that more distinctive node representation is provided for the downstream task.

Description

Technical Field

[0001] This invention relates to the technical field of graph structure data processing, specifically to a graph structure representation learning method and system based on the decoupling of conditional graphs and dual views. Background Technology

[0002] With the widespread application of complex and interconnected data such as biological networks, knowledge graphs, recommendation networks, communication networks, and industrial equipment networks, graph representation learning requires extracting effective features from nodes, edges, and their topological relationships, and transforming them into node representations for downstream tasks. Existing graph representation learning methods typically rely on explicit adjacency relationships in the original graph, using methods such as random walks, matrix factorization, graph neural networks, or contrastive learning for node feature propagation and structural representation learning. However, in real-world graph data, the original graph often suffers from sparse connections, insufficient local topological information, and unexpressed potential relationships. Representation learning based solely on the original graph view is insufficient to fully utilize the implicit structural information within the graph.

[0003] Existing graph augmentation or auxiliary graph construction methods typically generate new graph views through random perturbation, edge addition / removal, feature enhancement, or node similarity calculation. However, these generation processes often lack effective constraints on the overall structural characteristics and local structural differences of the original graph, leading to auxiliary structures that may deviate from the true topological distribution of the original graph or even introduce invalid or interfering connections. Furthermore, existing multi-view graph representation learning methods usually encode the original and augmented graph views separately and then directly concatenate, weightedly fuse, or simply align them. This coarse utilization of information between views makes it difficult to fully coordinate the relationships between original topological information, node semantic information, and potential supplementary structural information. Consequently, the final node representation suffers from insufficient information expression, inadequate structural discrimination ability, and limited adaptability to downstream tasks. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a graph structure representation learning method and system based on the decoupling of conditional graphs and dual views. It generates auxiliary graph views based on the statistical conditions of the original graph structure, and enhances the ability of node representation to comprehensively express explicit topological relationships and potential structural relationships by encoding and fusing the nodes of the original graph views and auxiliary graph views.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: One or more embodiments provide a graph structure representation learning method based on conditional graphs and dual-view decoupling, comprising the following steps: Obtain the graph data to be processed in the task scenario and construct the original graph view including the node set and edge set; Extract the node position representation of the nodes in the original graph view, and construct the structure condition vector of each node based on the node degree, average neighbor degree and graph density statistics based on the original graph view, thus forming the structure condition matrix; The node position representation and the structural condition matrix are input into the trained conditional diffusion network to generate an auxiliary node feature matrix; candidate neighbors are selected for each node based on the similarity between the auxiliary node features to construct an auxiliary graph view. The node position representations in the original graph view and the auxiliary graph view, the normalized adjacency matrix of the original graph and the normalized adjacency matrix of the auxiliary graph are input into the trained dual-view representation learning model to encode the dual-view nodes, resulting in node local embedding, original graph view structure embedding and auxiliary graph view structure embedding. The node local embedding, the original graph view structure embedding, and the auxiliary graph view structure embedding are fused to obtain the final node representation, and the downstream tasks of the task scenario are executed based on the final node representation.

[0006] One or more embodiments provide a graph structure representation learning system based on conditional graphs and dual-view decoupling, including: The original graph view construction module is configured to acquire the graph data to be processed in the task scenario and construct an original graph view including a set of nodes and a set of edges. The structure condition generation module is configured to extract the node position representation of the nodes in the original graph view, and construct the structure condition vector of each node based on the node degree, average neighbor degree and graph density statistics based on the original graph view, thus forming the structure condition matrix. The auxiliary graph view generation module is configured to input the node position representation and the structural condition matrix into a trained conditional diffusion network to generate an auxiliary node feature matrix; and to select candidate neighbors for each node based on the similarity between the auxiliary node features to construct the auxiliary graph view. The dual-view node encoding module is configured to input the node position representations in the original graph view and the auxiliary graph view, the normalized adjacency matrix of the original graph and the normalized adjacency matrix of the auxiliary graph into the trained dual-view representation learning model to encode the dual-view nodes, thereby obtaining the node local embedding, the original graph view structure embedding and the auxiliary graph view structure embedding. The fusion module is configured to fuse the local node embedding, the original graph view structure embedding, and the auxiliary graph view structure embedding to obtain the final node representation, and to execute the downstream tasks of the task scenario based on the final node representation.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the problem of insufficient node representation information due to sparse connections and unexpressed potential relationships in the original graph structure. It extracts node position representations from the original graph view and constructs a structure condition matrix by combining node degree, average neighbor degree, and graph density. This constrains the auxiliary graph generation process with both local node structural features and overall graph structural statistics. Furthermore, a trained conditional diffusion network generates an auxiliary node feature matrix, and candidate neighbors are selected based on the similarity of auxiliary node features to construct an auxiliary graph view, thus supplementing the original graph view with potential structural relationships. Simultaneously, the normalized adjacency matrix of the original graph and the normalized adjacency matrix of the auxiliary graph are input into a trained dual-view representation learning model for node encoding. This yields local node embeddings, original graph view structure embeddings, and auxiliary graph structure view embeddings, ensuring that the node representation simultaneously includes the node's local semantics, the original graph topology, and the supplementary structural information from the auxiliary graph. Finally, by fusing these three types of embeddings, the final node representation is obtained, improving the information completeness and structural expressiveness of the node representation, thereby providing more discriminative node representations for downstream tasks.

[0008] The advantages of the present invention, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0009] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0010] Figure 1 This is the overall algorithm flowchart of the graph structure representation learning method based on conditional graph and dual-view decoupling in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the dual-view training process for shared and private representation collaborative constraints in Embodiment 1 of the present invention. Detailed Implementation

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0012] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0013] It should be noted that the terminology used herein is for describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0014] To overcome the problems of insufficient utilization of topological information in a single graph, weak coupling between auxiliary graph generation and the original graph, mixed information in dual views, and insufficient modeling of higher-order structures in existing technologies, this invention provides a representation learning method based on conditional graph generation and dual-view decoupling. This method generates an auxiliary graph under the structural constraints of the original graph, and uses the original and auxiliary graphs to form a dual-view learning framework. Through shared-private representation decoupling, local-global joint modeling, and hard-sample-driven training, it achieves effective representation of node relationships, potential associations, and higher-order structural semantics in the graph. Specific embodiments are described below.

[0015] Example 1 In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 2 As shown, a graph structure representation learning method based on the decoupling of conditional graphs and dual views includes the following steps: Step 1: Obtain the graph data to be processed in the task scenario and construct the original graph view including the node set and edge set; Step 2: Extract the node position representation of the nodes in the original graph view, and construct the structure condition vector of each node based on the node degree, average neighbor degree and graph density statistics based on the original graph view, thus forming the structure condition matrix; Step 3: Input the node position representation and structural condition matrix into the trained conditional diffusion network to generate an auxiliary node feature matrix; select candidate neighbors for each node based on the similarity between the auxiliary node features to construct an auxiliary graph view; Step 4: Input the node position representations in the original graph view and the auxiliary graph view, the normalized adjacency matrix of the original graph and the normalized adjacency matrix of the auxiliary graph into the trained dual-view representation learning model to encode the dual-view nodes, and obtain the node local embedding, the original graph view structure embedding and the auxiliary graph view structure embedding. Step 5: Fuse the local node embedding, the original graph view structure embedding, and the auxiliary graph view structure embedding to obtain the final node representation, and execute the downstream tasks of the task scenario based on the final node representation.

[0016] This embodiment addresses the problem of insufficient node representation information due to sparse connections and unexpressed potential relationships in the original graph structure. It extracts node position representations from the original graph view and constructs a structure condition matrix by combining node degree, average neighbor degree, and graph density. This constrains the auxiliary graph generation process with both local node structural features and overall graph structural statistics. Furthermore, a trained conditional diffusion network generates an auxiliary node feature matrix, and candidate neighbors are selected based on the similarity of auxiliary node features to construct an auxiliary graph view, thus supplementing the original graph view with potential structural relationships. Simultaneously, the normalized adjacency matrix of the original graph and the normalized adjacency matrix of the auxiliary graph are input into a trained dual-view representation learning model for node encoding. This yields local node embeddings, original graph view structure embeddings, and auxiliary graph view structure embeddings, ensuring that the node representation simultaneously includes the node's local semantics, the original graph topology, and the supplementary structural information from the auxiliary graph. Finally, by fusing these three types of embeddings, the final node representation is obtained, improving the information completeness and structural expressiveness of the node representation, thereby providing more discriminative node representations for downstream tasks.

[0017] Graph representation learning methods are used to vectorize graph structure data. This involves converting information in a graph structure, such as nodes, edges, the connections between nodes, and the overall structural features, into low-dimensional continuous vectors that are easier for computers to process.

[0018] Taking graph G=(V,E) as an example, where V represents the set of nodes and E represents the set of edges, after graph representation learning, each node V in the graph... i An embedding vector h can be generated accordingly. i The embedding vector is used to characterize the structural position, connectivity, and potential semantic features of a node in the graph, and can be used as input to subsequent machine learning or prediction models to perform link prediction, node classification, graph classification, graph similarity calculation, or other graph-level analysis tasks.

[0019] For example, in a social network, each person can be viewed as a node, and the friendship relationships between people can be viewed as edges. The task of graph representation learning is to compress each person's position, relationships, and scope of influence in the social network into a vector. For example, the vector corresponding to user A is [0.23, 0.71, -0.14, 0.56].

[0020] In step 1, the graph data to be processed in the task scenario is obtained, and the corresponding graph data can be extracted for specific domains; Specifically, the task scenarios can be industrial equipment potential fault association identification scenarios, item or resource recommendation scenarios, communication network anomaly detection scenarios, etc. This embodiment mainly takes the industrial equipment potential fault association identification scenario as an example for illustration. By extracting more distinctive node representations, the accuracy of industrial equipment potential fault association identification is improved. In one specific implementation, for the field of potential fault association identification in industrial equipment, the graph data to be processed can be an equipment operation status diagram. Basic information of the industrial equipment, equipment component information, sensor acquisition information, operating parameter information, historical fault information, and maintenance record information can be acquired. One or more of the following—industrial equipment, equipment components, sensors, operating parameters, fault types, and maintenance measures—are constructed as nodes. Edges are then constructed based on the assembly relationships, acquisition relationships, parameter coupling relationships, historical fault association relationships, and the correspondence between faults and maintenance measures among the nodes to obtain the equipment operation status diagram of the industrial equipment. In step 1, the data to be processed may specifically include: equipment number, equipment type, equipment model, installation location, and operating time of industrial equipment; component number, equipment to which the equipment belongs, assembly relationship, and physical connection relationship of equipment parts; sensor type, installation location, sampling frequency, and monitoring object of sensors; one or more operating parameters of industrial equipment, including temperature, vibration, current, speed, pressure, and flow rate; fault type, fault occurrence time, fault component, fault level, and fault handling result of industrial equipment; maintenance time, maintenance object, maintenance measures, and maintenance result of industrial equipment; as well as the subordinate relationship between equipment and equipment parts, the assembly relationship between equipment parts, the acquisition relationship between equipment and sensors, the correspondence between sensors and operating parameters, the coupling relationship between operating parameters, the historical correlation between operating parameters and fault types, and the correspondence between fault types and maintenance measures.

[0021] By organizing the aforementioned industrial equipment data and their relationships into a graph structure, we can avoid separating basic equipment information, sensor data, operating parameters, and historical fault records, enabling subsequent models to analyze equipment status and fault relationships within a unified topology.

[0022] In this embodiment, potential fault association identification may include predicting whether there is a potential association between equipment components and fault types, predicting whether there is a potential association between abnormal operating parameters and fault types, or predicting whether there is a potential coupling relationship related to faults between different equipment components. For example, it may predict the probability of association between abnormal bearing vibration amplitude and bearing wear fault, or predict the probability of association between abnormal motor temperature and cooling system fault.

[0023] When the association probability of a candidate node pair is greater than a preset threshold, the candidate node pair is identified as a potential fault association, and the corresponding equipment component, operating parameter, or fault type is output as an object to be investigated; when the association probability is not greater than the preset threshold, the candidate node pair is not identified as a potential fault association.

[0024] The nodes in the equipment operation status diagram include one or more of the following: industrial equipment nodes, sensor nodes, operating parameter nodes, fault type nodes, and maintenance record nodes; edges are used to characterize the acquisition relationship between equipment and sensors, the coupling relationship between sensor parameters, the assembly relationship between equipment components, the correlation between operating parameters and fault types, or the correspondence between equipment faults and maintenance measures.

[0025] For example, for motors, bearings, pumps, or reducers in a production line, sensor data such as temperature, vibration, current, speed, and pressure can be used as node attributes, and the mechanical connections or signal transmission relationships between equipment components can be used as edges, thus forming a graph-structured data for fault diagnosis or remaining life prediction. This approach enables subsequent models to comprehensively utilize equipment topology and operating status characteristics, improving the ability to identify abnormal states, potential faults, or degradation trends of critical components.

[0026] In another implementation, in a recommendation scenario, the graph data to be processed can be a user-item interaction graph. The nodes in the graph can include user nodes, item nodes, user interest tag nodes, item attribute nodes, and historical behavior nodes. Edges are used to represent interactive relationships between users and items, such as clicks, browsing, favorites, purchases, ratings, or dwell time. By constructing the above graph structure, it is possible to represent user preferences, item relationships, and the characteristics of user interests changing over time, thereby providing a graph structure data foundation for subsequent item recommendations, interest predictions, or similar user identification.

[0027] In the field of recommendation, link prediction can predict whether there is a potential interaction relationship between user nodes and item nodes, such as predicting whether a user is likely to click, browse, favorite, purchase, or rate an item. In step 1, we first obtain data that can form a graph from specific business or technical scenarios, and then organize this data into nodes and edges to form an original graph structure that can be processed by the model; the graph data can be different for different business or technical scenarios. In the field of potential fault association identification in industrial equipment, the acquired graph data to be processed may include: equipment data, sensor data, operating parameter data, fault type data, maintenance record data, as well as the acquisition relationship between equipment and sensors, the assembly relationship between equipment components, and the association relationship between operating parameters and fault types.

[0028] In the recommendation field, the graph data to be processed can include: user data, item data, user interest tag data, item attribute data, historical behavior data, and interaction data between users and items such as clicks, browsing, favorites, purchases, ratings, and dwell time.

[0029] Furthermore, the input graph data is acquired and represented as the original graph. ,in, Represents a set of nodes. Indicates the total number of nodes; Denotes the set of edges. Representing the total number of edges: For an undirected graph, if ,but They are considered to be the same connection relationship; for directed graphs, edge direction information is preserved.

[0030] In some embodiments, to facilitate subsequent tensor quantization calculations, the node identifiers in the original graph view are first renumbered with consecutive integers. Let the node renumbering be mapped as follows: Based on this, the renumbered graph is obtained; the edge set is converted into an edge index matrix. The m-th edge corresponds to the edge index. , Simultaneously construct the adjacency matrix of the original graph view. ,when hour, ,otherwise For an undirected graph, the adjacency matrix... .

[0031] In the scenario of identifying potential fault associations in industrial equipment, the node set Includes one or more of the following: industrial equipment nodes, equipment component nodes, sensor nodes, operating parameter nodes, fault type nodes, and maintenance measure nodes; edge set. This is used to represent the assembly relationships between equipment components, the data acquisition relationships between equipment and sensors, the correspondence between sensors and operating parameters, the coupling relationships between operating parameters, and the historical association relationships between operating parameters, equipment components, and fault types. The resulting adjacency matrix... It can uniformly represent the dispersed physical connections, signal transmissions, and historical fault association information in industrial equipment systems, providing a structured data foundation for subsequent identification of potential fault associations.

[0032] The above implementation method of this embodiment addresses the problems of sparse connection of existing equipment operation status diagrams and incomplete expression of fault association information caused by the dispersed sources of industrial equipment operation data, complex equipment topology relationships, limited historical fault samples, and the lack of explicit recording of potential fault association relationships. First, the industrial equipment, equipment components, sensors, operating parameters, fault types, and maintenance measures are organized into an original diagram view of industrial equipment operation status to uniformly represent the physical connection relationship, data acquisition relationship, operating parameter coupling relationship, and historical fault association relationship of the equipment.

[0033] In step 2, the original graph view is used as input to extract the positional representations of the nodes, and a structure condition matrix is ​​constructed based on the topological statistics in the original graph view. This structure condition matrix serves as the structural constraint input for the subsequent conditional diffusion network, ensuring that the generated auxiliary node features retain the local connectivity features and global structural features of the original graph view. Specifically, the location and structural feature extraction module is constructed to encode the topological positional relationships of nodes in the original graph view, resulting in a node position representation matrix. Each row in the node position representation matrix corresponds to the position representation of a node, which is used to characterize the relative topological positional relationship of the corresponding node in the original graph view. In the scenario of identifying potential fault associations in industrial equipment, node location representations are used to characterize the relative structural positions of equipment, components, sensors, operating parameters, and fault types within the equipment's operational status diagram. For example, sensor nodes and operating parameter nodes connected to the same equipment component may have similar topological location representations, while equipment component nodes with historical associations to the same fault type can form corresponding association patterns in the structural representation space. By extracting node location representations, the assembly hierarchy, monitoring relationships, and fault association structures within the equipment system can be preserved, avoiding isolated judgments based solely on individual sensor values.

[0034] In step 2, the structure condition matrix is ​​used to describe the structure condition information corresponding to each node, and is used to characterize the local structure or the whole graph structure. In this embodiment, the structure condition information includes node degree, average neighbor degree, and graph density. Specifically, the node degree is the number of connected nodes of node i in the original graph view; For any node i, define its degree. The number of connections for node i in the original graph view is expressed by the formula: ; in, The degree of node i represents the local connectivity strength of the node. The elements in the adjacency matrix of the original graph view represent the connection relationship between node i and node j. Specifically, the average degree of node i's neighbors is the average degree of all neighbor nodes in node i's neighbor set; Define the set of neighbors of node i. for: ; Define the average degree of the neighbors of node i as: ; in, Let be the average degree of the neighbors of node i, which is used to characterize the average structural activity of the node's neighborhood; Specifically, define the graph density of the original graph view. The ratio of the actual number of edges in the original graph view to the theoretical maximum number of edges that can exist in the original graph view is expressed by the formula: ; in, Graph density is used to characterize the overall density of the original graph view. This indicates the number of edges in the original graph view; For a directed graph, it can be written as follows: ; Based on the above statistics, a structural condition vector is constructed for each node. : ; Construct a structure condition matrix based on the structure condition vectors of all nodes. : ; Structural condition matrix As a graph structure constraint input for subsequent conditional diffusion networks, it guides the generation process of auxiliary node features.

[0035] In the scenario of identifying potential fault associations in industrial equipment, node degree Used to characterize the number of known associations between a corresponding device object and other device objects; average neighbor degree. Used to characterize the average association activity level of the device subsystem to which the node belongs; graph density This is used to characterize the overall connectivity density of the entire equipment's operational status diagram. For example, the node degree of a device component node can reflect the number of connected sensors, operating parameters, and historical fault types, while the neighbor average degree can reflect the structural complexity of its subsystem. These local and global structural statistics are constructed into a structural condition matrix. This allows the generation process of subsequent auxiliary node features to be constrained by the actual topology of the device and the historical association distribution, reducing the risk of generating invalid associations that are unrelated to the device structure.

[0036] Optionally, node attributes can be used as input data for the conditional diffusion network; if the original graph view contains node attributes, the node attribute matrix can be further denoted as... ,in For attribute dimensions; if there are no external attributes, only the position representation matrix will be used in subsequent training. and structural condition matrix As input data.

[0037] In this embodiment, the conditional diffusion network is implemented using an existing conditional diffusion generation network, which is used to generate auxiliary node features based on the structural condition information of the nodes. Specifically, the conditional diffusion network uses random noise state as the starting point of the reverse generation process, and inputs the structural condition vector of the node in each reverse diffusion step, so that the denoising process is constrained by structural statistical information such as node degree, average neighbor degree, and graph density, thereby generating auxiliary node features that match the structural characteristics of the original graph.

[0038] In step 3, the node position representation and the structural condition matrix are input into the trained conditional diffusion network to generate an auxiliary node feature matrix. Based on the similarity between the auxiliary node features, candidate neighbors are selected for each node to construct an auxiliary graph view, including the following steps: Step 31: For each node, sample a random noise vector from a standard Gaussian distribution as the node's noise state at the initial diffusion step T of the reverse generation process. ); ;in, This represents a matrix with a mean of 0 and a covariance equal to the identity matrix. The distribution is Gaussian, where T represents the total number of diffusion steps. This represents the initial Gaussian noise vector of node i at the Tth diffusion step, serving as the starting state of the reverse generation process. In the diffusion generation model, the reverse generation process starts from the Tth step, and the Tth step usually corresponds to a state close to pure noise. Therefore, this Gaussian noise vector is the initial noise state of node i at the Tth diffusion step. Step 32: Based on the noise state at the initial diffusion step T, proceed according to the sampling step... arrive The sequential execution conditional backsampling, for the first One diffusion step, the current noise state diffusion step and the structural condition vector of node i Input the trained conditional diffusion network to obtain the estimated state of node i in the previous diffusion step. : ; in, This represents the processing procedure of the trained conditional diffusion network. This represents the network parameters after training. This represents the previous diffusion step output by the conditional denoising network. The state estimate.

[0039] Step 33: Based on the estimated state from the previous diffusion step, iteratively update the node states according to the conditional backsampling rule to obtain the auxiliary node features; the iterative form is as follows: ; in, This represents Gaussian noise, which is the random perturbation noise added at each step during the backsampling process; Indicates the first Noise scheduling parameters corresponding to each diffusion step; After the iteration ends, we get the first... Auxiliary features of each node Then, all nodes are stacked in order of their node numbers to obtain the auxiliary node feature matrix. : ; in, Represents the auxiliary node characteristics of the i-th node. This represents the feature dimension of the auxiliary node.

[0040] Auxiliary node feature matrix Representation of the original position They have the same node correspondence and feature dimensions, but their feature distribution has been reconstructed through structural condition constraints, which can serve as an auxiliary representation basis for subsequent difficult negative sample mining and multi-view alignment learning.

[0041] For identifying potential fault associations in industrial equipment, auxiliary node features can characterize the associations between equipment components and fault types, operating parameters and fault types, and different equipment components, while preserving equipment assembly relationships, sensor acquisition relationships, and historical fault association distributions. This avoids relying solely on single sensor parameters or local historical records for fault association judgment.

[0042] Step 34: Calculate the similarity between the auxiliary node features of any two nodes, and then sort them according to the similarity. Sort the nodes from highest to lowest similarity and select the top K nodes with the highest similarity to form the candidate neighbor set for node i. Construct an auxiliary adjacency matrix based on the candidate neighbor set. ; In obtaining the feature matrix of auxiliary nodes Then, the auxiliary graph structure is further reconstructed based on the feature similarity between nodes, where, Represents a node With nodes The similarity matrix between auxiliary node features is defined as follows: ; To avoid self-connection interference, let It does not participate in adjacency reconstruction. For each node... Select the one with the highest similarity Each node constitutes a candidate neighbor set. ; In one embodiment, It can be a fixed positive integer.

[0043] Construct an auxiliary adjacency matrix based on the candidate neighbor set. ,in: ; Step 35: Determine the set of auxiliary edges based on the auxiliary adjacency matrix. Then construct auxiliary graph view : ; Among them, the auxiliary edge set It is composed of auxiliary adjacency matrix The node pairs with a value of 1 are determined; the auxiliary graph view and the original graph view share the same set of nodes. However, its edge connections are induced by auxiliary node features generated by a pre-trained conditional diffusion generator network, which is used to characterize the potential relationships between nodes from another perspective.

[0044] In the scenario of identifying potential fault associations in industrial equipment, auxiliary edges can represent potential coupling relationships between equipment components and fault types, operating parameters and fault types, or equipment components that are not explicitly recorded in the original equipment operating status diagram. By constructing auxiliary graph views, potential fault association structures can be supplemented without changing the original equipment topology and historical association relationships. This can supplement the potential association relationships between equipment, sensors, operating parameters, and fault types that are not explicitly recorded in the original graph views, providing a structural basis for subsequent dual-view node encoding and candidate fault relationship identification.

[0045] Further technical solutions also include the training process of the conditional diffusion network, which includes the following steps: Step S301: Obtain the node position representation matrix: based on the training graph The nodes in the training graph are encoded to obtain a node position representation matrix: ; Where N represents the number of nodes. This represents the position of the i-th node. Node position representation Used to characterize the relative positional relationships of nodes in the training graph topology, and as clean samples during the training of the conditional diffusion network. In other words, in the training of the conditional diffusion network, It represents the target node that the network is expected to eventually recover or generate; Among them, training images The edge-leaving partition originates from the original graph. Let the original graph be: ; From the original edge set A subset of real edges is reserved as the test positive sample set, and these edges are removed from the original graph to obtain the training edge set, thus yielding the training graph. : ; Step S302: Process the training image Statistical node-level and graph-level structural features: for training graphs For any node i in the array, calculate its degree. And determine the neighbor set of node i; calculate the average neighbor degree of node i based on the node degree of each neighbor node in the neighbor set; calculate the graph density δ based on the number of nodes and edges in the training graph; construct the structure condition vector of node i based on the node degree, average neighbor degree, and graph density; wherein, the calculation methods of node degree, average neighbor degree, and graph density are the same as those in step 2 above; for the training graph Statistical node-level and graph-level structural features.

[0046] If the original graph also contains node attributes, then the node attribute matrix can be further denoted as... ,in For attribute dimensions; if there are no external attributes, only the position representation matrix will be used in subsequent training. and structural condition matrix This serves as the input basis. Therefore, the final output of this step is the training image. Test sample set Node position representation matrix Edge index matrix Adjacency matrix and structural condition matrix .

[0047] In this step, the obtained training image Node position representation matrix and structural condition matrix Using the input as input, a conditional diffusion network is constructed. The goal of the conditional diffusion network is to learn the conditional generation distribution of node position representations while preserving the statistical regularities of the local structure of the original graph, and to generate auxiliary node feature matrices accordingly. Further reconstruction yields an auxiliary graph. .

[0048] Step S303: Set the total number of diffusion steps And noise scheduling parameters, and construct forward-noise-added samples. To achieve conditional generation, first establish the total number of diffusion steps as... The forward noise addition process.

[0049] Record No. The node at the th The noisy state under each diffusion step is ,in Let the noise scheduling sequence be... and define: , Indicates the first Noise scheduling parameters for each diffusion step, This represents the signal retention coefficient of the diffusion step. Indicates as of the date The cumulative signal retention coefficient of each diffusion step. This is the index variable in the cumulative product. Represented by the aforementioned node position. As a clean sample The forward noise addition process of each diffusion step is represented as follows: ,in, Indicates the first Gaussian noise is sampled in a diffusion step. This allows the original node position representation to be progressively mapped to the noisy state space, thus providing training samples for subsequent conditional denoising learning.

[0050] Step S304: During the reverse generation process, the construction parameters are as follows: Conditional denoising network The original position features of nodes are recovered from the forward-noiseed samples by using a conditional denoising network. The conditional denoising network is trained with the goal of minimizing the mean square error between the estimated value and the true position representation, and a trained conditional denoising network is obtained. Specifically, to ensure the generated results are constrained by the graph structure, diffusion step embedding and structural condition embedding are introduced. Let the diffusion step... Time embedding as Node condition vector The input to the denoising network can be represented as the fusion result of the noise state and conditional information, and the output is an estimate of the original position representation. ,Right now: ; During training, the node set is randomly sampled, and the sampling is performed on the diffusion step set. The step size is randomly selected to minimize the mean square error between the estimated value and the true location representation, which serves as the training objective of the conditional diffusion network. ; Through the above training, the conditional diffusion network learns the mapping relationship from random noise to node features constrained by structural conditions, thus generating auxiliary node representations while preserving the statistical properties of the original graph. In the generation phase, an initial noise vector is first sampled from a standard Gaussian distribution for each node. Then according to from arrive The sequential execution of conditional reverse sampling yields auxiliary node features.

[0051] In the above implementation of this embodiment, to address the problems of sparse graph structure, numerous noisy edges, and difficulty in explicitly modeling potential higher-order relationships, an original graph view is constructed. A structural condition vector is then built based on node position representation combined with node degree, average neighbor degree, and graph density. Further, a conditional diffusion network is used to generate auxiliary node features under structural constraints. An auxiliary graph view is then constructed based on the similarity between these auxiliary node features, thereby introducing supplementary structural information beyond the original graph. By introducing explicit statistical constraints during the structure generation process, the generated auxiliary graph maintains consistency with the original graph's structural distribution, thus reducing the risk of introducing noisy edges in unconstrained graph enhancement. Simultaneously, the auxiliary graph view supplements the modeling of potential higher-order relationships not explicitly expressed in the original graph, improving the ability to express complex structural relationships. Furthermore, it enhances the completeness of graph structural information when the original graph structure is sparse or lacks local connectivity, thereby improving the overall expressive quality of graph representation learning.

[0052] Specifically, addressing the issues of sparse connectivity, incomplete historical fault association records, and difficulty in explicitly expressing potential high-order coupling relationships between equipment components in industrial equipment operation status graphs, this paper constructs structural condition vectors based on node location representation, node degree, average neighbor degree, and graph density. Then, it utilizes a conditional diffusion network to generate auxiliary node features under structural condition constraints, thereby constructing an auxiliary graph view of potential equipment associations. This approach, while maintaining the original equipment topology and historical fault association distribution, supplements the unrecorded potential associations between equipment components, operating parameters, and fault types, reducing the risk of introducing invalid connections through unconstrained graph enhancement.

[0053] Based on the original and auxiliary graph views described above, a dual-view representation learning model is further constructed. In the scenario of identifying potential fault associations in industrial equipment, the original graph view is used to retain the existing explicit topological relationships and historical association information in the industrial equipment operation status graph, while the auxiliary graph view is used to supplement the potential association structure induced by node features. By jointly encoding the two complementary views, the ability of node representation to express complex equipment association relationships can be improved.

[0054] Furthermore, the dual-view representation learning model in step 4 is as follows: Figure 1 As shown, it includes a local encoder configured in parallel, an original image view encoding branch and an auxiliary image view encoding branch, as well as a projection fusion head; The local encoding module is used to extract the semantic information and state features of the nodes themselves; the original graph view encoding branch and the auxiliary graph view encoding branch are used to model the structural relationships in the original graph view and the auxiliary graph view, respectively; Both the original graph view encoding branch and the auxiliary graph view encoding branch adopt a concatenated encoding structure, including a hypergraph global encoder and a GCN encoder. The hypergraph global encoder is used to model higher-order relations based on the node-hyperedge association in the view and obtain the initial value of the higher-order node representation containing higher-order semantic association. The GCN encoder takes the initial value of the higher-order node representation as input and combines it with the normalized adjacency matrix of the corresponding view to perform structure propagation and obtain the structure embedding of the corresponding view. The projection fusion head is used to fuse local node embeddings, original graph-view structure embeddings, and auxiliary graph-view structure embeddings to obtain the final node representation. Specifically, the fusion method can employ at least one of the following: stitching, weighted summation, attention fusion, or multilayer perceptron mapping.

[0055] In a specific implementation, in the original graph view encoding branch and the auxiliary graph view encoding branch, the hypergraph global encoder is used to perform high-order relationship modeling based on the node-hyperedge association relationship in the corresponding view. Specifically: first, an association matrix between nodes and hyperedges is constructed according to the node representation. Through the process of node to hyperedge, hyperedge update, and hyperedge to node propagation, high-order group relationship aggregation is realized to obtain the initial value of the high-order node representation. For the original graph view and auxiliary graph view of industrial equipment operation, the hypergraph global encoder performs aggregation modeling of high-order group relationships between multiple equipment components, sensors, operating parameters, or fault types. In a specific implementation, the GCN encoder is used as the initial value of the high-order node representation output by the hypergraph global encoder as the node feature input, and combines the normalized adjacency matrix of the corresponding view to perform graph convolution propagation. By utilizing the explicit connection relationships in the original graph view or auxiliary graph view, topological constraints and local structure optimization are performed on the high-order node representation, and finally the structure embedding under the corresponding view is obtained. The dual-view representation learning model constructed in this embodiment learns the representation of nodes from three levels: node semantics, explicit topological relationships, and latent structural relationships, through a local encoding module, an original graph view encoding branch, and an auxiliary graph view encoding branch. The original graph view GCN encoder outputs an original graph view structure embedding to characterize the known assembly relationships, sensor acquisition relationships, and historical fault correlations between equipment components. The auxiliary graph view GCN encoder outputs an auxiliary graph view structure embedding to characterize the equipment component-fault type, operating parameter-fault type, and potential correlations between equipment components, supplemented by the auxiliary graph. For industrial equipment operation scenarios, the original graph view structure embedding is used to characterize the known assembly relationships, sensor acquisition relationships, and historical fault correlations between equipment components; the auxiliary graph view structure embedding is used to characterize the equipment component-fault type, operating parameter-fault type, and potential correlations between different equipment components, supplemented by the auxiliary graph.

[0056] The final node representation output by the projection fusion head includes the node's own local semantics, the original graph's true topology information, and the auxiliary graph's potential structural information. For industrial equipment operation scenarios, the final node representation includes the node's own state information, the original equipment topology and historical fault information, as well as potential fault association information in the auxiliary graph. It can be used to identify potential associations between equipment components and fault types, operating parameters and fault types, or different equipment components.

[0057] Furthermore, the dual-view representation learning model also includes a sample construction and sampling module, as well as a two-stage training unit; The sample construction and sampling module is used to construct positive sample node pairs based on random walks of the original graph view, mine difficult negative samples based on auxiliary graph views, and combine random negative sampling to form a node pair training set; and further generate batch subgraph structures for mini-batch training, providing supervised labels and structural inputs for dual-view representation learning; In the scenario of identifying potential fault associations in industrial equipment, positive samples are used to characterize known direct or higher-order associations between equipment components, sensors, operating parameters, and fault types; hard negative samples are used to characterize node pairs that have high potential associations in the auxiliary graph but have no known connections in the original equipment operating state graph, thereby improving the model's ability to distinguish between real fault associations and potential confusing relationships.

[0058] A two-stage training unit is configured to optimize the dual-view representation learning model using a two-stage training strategy; In the first stage, the connection relationship supervision and discrimination is performed based on the positive and negative sample node pairs output by the sample construction and sampling module, and the synergy between node local embedding, original graph view structure embedding and auxiliary graph view structure embedding is constrained so that the local representation can be jointly guided by the high-order structure information of the original graph and the high-order structure information of the auxiliary graph. The second phase involves performing cross-view alignment, shared information learning, and private information decoupling across the entire graph. In the scenario of identifying potential fault associations in industrial equipment, the first stage of training enables node representations to be guided by both known equipment topology information and potential fault association information. The second stage of training enables the common semantics of the same equipment object to remain consistent in the original graph view and the auxiliary graph view, and reduces the repetitive expression between the unique structural information of the two views, thereby obtaining node representations with stronger fault association discrimination capabilities.

[0059] After the model training is completed, the sample construction and sampling module and the two-stage training unit do not participate in inference; in the application stage, the trained local encoder, the original graph-view encoding branch, the auxiliary graph-view encoding branch and the projection fusion head are called to obtain the final node representation, which is used for the identification of potential fault associations between equipment components and fault types, operating parameters and fault types, or different equipment components.

[0060] In some embodiments, the step of training the dual-view representation learning model includes the following: Step S401: Divide the acquired training graph into data and construct a training graph; Specifically, let the edge margin ratio be... Then from the original edge set Selected from The set of real edges is used as the set of positive samples to be reserved. , recorded as After removing this portion of the edges from the original graph, the training graph is obtained. , where the symbol " "This is a removal operation; In a preferred embodiment, to ensure the connectivity of the training graph, the set of non-deletable edges is first calculated based on the original graph. The remaining edges are the set of edges that can be deleted; for an undirected graph, the set of edges that cannot be deleted is the set of edges of the minimum spanning tree of the original graph, and for a directed graph, the set of edges that cannot be deleted is the set of edges of the minimum branch tree.

[0061] Based on this, positive samples are selected from the set of deletable edges to ensure that the training graph still maintains its connected structure after the edges are deleted.

[0062] From the non-edge set of the original graph view The same number of node pairs as the number of positive samples are randomly sampled to form a set of negative samples, which constitutes the test negative samples for the link prediction task. Specifically, the set of non-edges is defined as: ; From non-edge set Randomly sample the same number of node pairs as the number of positive samples to form the set of retained negative samples. Based on this, a test sample set is formed: ; Among them, when hour, ;when hour, , This represents the label of the node pair (i,j).

[0063] Therefore, the training graph with edges removed is used during the training phase. The test phase uses a sample set. Complete the link prediction assessment.

[0064] Furthermore, after completing the graph partitioning, an initial position representation is constructed for each node in the training graph. The training graph is then encoded using a set position encoding learning module to obtain a node position representation matrix, with each node corresponding to a position representation vector. Specifically, let the position encoding dimension be... Learning module through location representation For training images Encode the nodes to obtain a matrix representing their positions. , This represents the position of the i-th node.

[0065] Step S402: For the original graph view in the training graph, perform the aforementioned steps 2 and 3 to construct an auxiliary graph view; Specifically, the node position representations of the nodes in the original graph view are extracted, and the structural condition vector of each node is constructed based on the node degree, average neighbor degree and graph density statistics based on the original graph view, forming a structural condition matrix; The node position representation and the structural condition matrix are input into the trained conditional diffusion network to generate an auxiliary node feature matrix; candidate neighbors are selected for each node based on the similarity between the auxiliary node features to construct an auxiliary graph view. Step S403: Construct positive sample node pairs by performing a random walk mechanism based on the original graph view in the training graph, and mine difficult negative sample node pairs based on the auxiliary graph structure. At the same time, combine the random negative sampling strategy to form a node pair training set. This step is implemented in the sample construction and sampling module. In this step, the training graph is used. Auxiliary diagram Node position representation matrix and auxiliary node feature matrix As input, a node pair training set is constructed for the first stage of supervised discriminative learning, and a structural relation matrix corresponding to the original graph and the auxiliary graph is established simultaneously. The node pair training set consists of a set of positive sample node pairs and a set of negative sample node pairs, where positive samples are used to represent local co-occurrence relationships in the original graph, and negative samples are used to highlight potential confusing connections revealed by the auxiliary graph, thereby enhancing the model's discriminative ability.

[0066] Step S4031, Construction of positive sample node pairs: Perform a biased random walk operation on the original graph. For any random walk sequence, take the starting node in the sequence as the target node and the other visited nodes in the same sequence as the context nodes, and then construct a set of positive sample node pairs. Specifically, in the original view Perform a biased random walk on the node. Let the number of random walks for each node be . The distance traveled each time was Returns bias parameters External bias parameters For any random walk sequence, let the nth... Step access node is ,in Indicates the first In the next random walk, given the previous node... and the current node In this case, the next node The transition probability can be expressed as: ; in, Represents a node Neighbor set, bias factor Defined as: ; in, This represents the return bias parameter, used to adjust the relative transfer weights from the current node back to the previous node. The larger the value, the weaker the tendency to return to the previous node; This represents the outward bias parameter, used to adjust the relative transfer weight when the candidate node is neither the previous node nor adjacent to the previous node; The smaller the value, the more the random walk tends to explore outwards from the current local neighborhood.

[0067] Therefore, a set of random walk sequences can be obtained from the original graph view. For any walk sequence Starting node As the target node, other visiting nodes in the same sequence are used as context nodes to construct a set of positive sample node pairs: Among them, tags This indicates that the corresponding node pair is a positive sample.

[0068] In this embodiment, the positive samples obtained by the above method not only retain the direct adjacency relationship in the original graph, but also introduce higher-order co-occurrence information on the random walk path, which can more stably characterize the local topological correlation of nodes.

[0069] Step S4032, Construction of difficult negative samples: based on auxiliary graph view Structure, selecting node pairs to construct difficult negative samples; Specifically, difficult negative samples are selected by prioritizing node pairs that have high potential correlation in the auxiliary graph but no real connection in the original training graph, thus making them more difficult to distinguish. To this end, a first-class candidate set of difficult negative samples is defined. for: ; From the set A preset number of node pairs are selected to form the first part of the difficult negative samples.

[0070] Let the proportion of difficult negative samples be... The number of positive samples is The number of difficult negative sample targets is: If set If the number of candidate node pairs in the auxiliary graph is insufficient, then the second-order topological relationships in the auxiliary graph are used to supplement the difficult negative samples. Specifically, a second-class difficult negative sample candidate set is defined. for: ; For candidate node pairs Nodes with higher connectivity in the auxiliary graph are prioritized to improve the confusion of negative samples. Let nodes be... The node degree in the auxiliary graph is Then according to node degree Candidate node pairs are selected sequentially after sorting from largest to smallest. If the number of difficult negative samples is still insufficient, node pairs are randomly sampled from the non-edge set of the original training graph as supplementary negative samples to form a random negative sample set. The final set of negative samples is obtained as follows: ; in, Indicates a set and The selected set of difficult negative samples, with corresponding labels as follows: The positive sample set With negative sample set Merge and randomly shuffle the nodes to obtain the node pairs in the training set. : ; in, For node pairs Supervisory labels; when Time represents a positive sample node pair, when The node pair represents the negative sample node pair. This node pair training set will be used in the next step for supervised discriminative learning, so that the local representation learning can be subject to the dual constraints of the local co-occurrence relations of the original graph and the difficult negative samples of the auxiliary graph.

[0071] Step S404: Normalize the adjacency matrices of the original graph view and the corresponding auxiliary graph view in the training graph, including adding an identity matrix to the adjacency matrix to introduce self-loops, and performing symmetric normalization based on the node degree matrix to obtain the normalized adjacency matrix. and And the batch subgraph structure corresponding to the training batch; Specifically, normalized adjacency inputs are constructed for two types of graphs (i.e., the original graph view and the auxiliary graph view). Let the adjacency matrix of the original graph view in the training graph be... The adjacency matrix of the auxiliary graph view is Add the identity matrix to each adjacency matrix. After introducing self-loops, the node degree matrices can be obtained as follows: ; ; Furthermore, construct the normalized adjacency matrix: ; ; in, Used to depict normalized structural relationships under the original diagram view. This is used to characterize the normalized structural relationships under the auxiliary graph view. The two types of normalized adjacency matrices can serve as the input basis for subsequent global structure encoding, higher-order relationship propagation, and dual-view collaborative modeling.

[0072] In a preferred embodiment, to accommodate small-batch training, a unique set of nodes can be further extracted from the set of node pairs in the current batch. Let the set of node pairs in a certain batch be denoted as . Then the set of unique nodes involved is denoted as: ; Furthermore, from and Extraction and The corresponding submatrices are denoted as: ; ; The batch subgraph structure input, along with the node position representations corresponding to that batch, auxiliary node features, and node pair labels, is fed into the next step to complete local discriminative learning and dual-view high-order structure modeling. Thus, this step ultimately outputs a node pair training set. Normalized adjacency matrix of the original graph Auxiliary graph normalized adjacency matrix And its corresponding batch subgraph structure input.

[0073] To address the shortcomings of existing training mechanisms, such as insufficient representation of positive sample structure information, reliance on random sampling for negative samples leading to limited discrimination difficulty, and inconsistent input scales in dual-view structures making them unsuitable for small-batch training, this embodiment constructs positive sample node pairs based on biased random walks of the original graph view to introduce direct adjacency relationships and higher-order co-occurrence information. It also mines difficult negative samples with potential associations but no real connections in the original graph based on the auxiliary graph view, supplementing them with second-order topological relationships and random negative sampling when the number is insufficient, thereby improving the discrimination difficulty and coverage of training samples. Simultaneously, self-loops are added to the adjacency matrices of the original and auxiliary graph views and normalized, and batch subgraph structure inputs are constructed based on the current batch of node pairs. This enhances the model's ability to distinguish between real connections and potential confusion relationships, improves the stability of dual-view structure propagation, and reduces computational overhead during small-batch training.

[0074] Step S405, First stage training: The node connection relationship is learned by the supervised discriminative loss constraint model of positive and negative sample node pairs, and the dual-view representation learning model is jointly optimized in the first stage by the collaborative constraint between the local representation and the dual-view higher-order representation. The nodes obtained from the above steps are used in the training set. Normalized adjacency matrix of the original graph Auxiliary graph normalized adjacency matrix Node position representation matrix and optional node attribute matrix The first phase of training is performed using the input. The goals of the first phase of training are: on the one hand, to complete supervised connection discrimination for node pairs; on the other hand, to establish a consistent representation relationship between the local representation of the node and the higher-order representation of the original graph and the higher-order representation of the auxiliary graph, thereby forming a dual-view modeling mechanism that coordinates local and global aspects.

[0075] Furthermore, the first phase of training includes the following steps: Step S4051: Based on the batch subgraph structure, in each round of mini-batch training, select a batch of node pairs from the node pair training set and extract the node position representation submatrix corresponding to the current batch. Original graph subadjacency matrix and auxiliary graph subadjacency matrix ; Specifically, for any batch First, extract the set of its corresponding unique nodes. Based on this, a submatrix representing the node positions of this batch is formed. Original graph subadjacency matrix and auxiliary graph subadjacency matrix .

[0076] Optionally, if node attributes are available, an optional node attribute submatrix can be constructed. ; Step S4052, Local Representation Construction: Submatrix Based on Node Position Representation The data is input to the corresponding local encoder, which performs fine-grained feature encoding on the node's own features and its local neighborhood structure to obtain the node's local embedding. For each node in the batch Through a local encoder, let the position encoding mapping be... Attribute encoding mapping is Then the node local embedding representation of node i is defined as: ; in, This represents the node local embedding representation of the i-th node. This indicates vector concatenation.

[0077] Local embedding primarily preserves the node's own positional and attribute semantic information, which is used for subsequent node pair discrimination and local-global collaborative constraints.

[0078] Step S4053: For the original graph view and the auxiliary graph view, perform hypergraph high-order relation modeling and GCN structure propagation through the corresponding branch hypergraph encoder and GCN encoder to obtain the original graph view structure embedding and the auxiliary graph view structure embedding.

[0079] Step S4053 constructs the structural embedding under the original view and the auxiliary view respectively. The specific process is described below; Step S40531: Concatenate the position information and attribute features of each node to obtain the global input vector of the node. When a node does not have attribute characteristics, only the position of the node is represented. This serves as its global input vector. The global input vectors of each node within the batch are stacked to obtain the global input matrix. ; Specifically, define the global input vector of the node. : ; in, Indicates feature concatenation operation; Represents a node The node position is represented as follows: Represents a node The corresponding attribute characteristics.

[0080] Step S40532: For the original graph view, convert the global input matrix... Node projection representation is obtained through node projection mapping. ; Node-based projection representation and the hyperedge correlation parameter matrix Construct a node-hyperedge association matrix This characterizes the strength of the association between each node and the preset hyperedge.

[0081] Specifically, for the original graph view, the node projection representation is first obtained through node projection mapping: ; in, This represents the projection of nodes from the original view.

[0082] Based on this, construct the node-hyperedge association matrix: ; in, This represents the hyperedge association parameter matrix under the original graph view. Indicates the preset number of superedges. The Rows represent nodes The strength of the association with each hyperedge.

[0083] Step S40533: Model higher-order hypergraph relations based on node-hyperedge incidence matrix: Based on the node-hyperedge incidence matrix... Aggregate the node projection representations to obtain the hyperedge representation. Then, a nonlinear transformation is performed on the hyperedge representation using the hyperedge update mapping to obtain the updated hyperedge representation. Subsequently, the updated hyperedge representation is propagated back to the node side to obtain the initial values ​​of the higher-order node representations under the original graph view. ; Specifically, based on the node-hyperedge association matrix First, calculate the hyperedge representation: ; Then, the updated hyperedge representation is obtained by updating the mapping through the hyperedge: ; in, Update the parameter matrix for the hyperedge in the original graph view. Represents a non-linear activation function; Subsequently, the updated hyperedge representation is propagated back to the node side to obtain the initial values ​​of the higher-order node representation under the original graph view: ; in, This represents the initial value of the higher-order node representation obtained after the original graph view is processed by the hypergraph global encoder for node-to-hyperedge, hyperedge update, and hyperedge-backward node propagation. It is used as the node feature input for the subsequent GCN encoder.

[0084] Similarly, for auxiliary graph views, the node-hyperedge association matrix under the auxiliary graph view is constructed using independent parameters. Following the process of node-to-hyperedge, hyperedge update, and hyperedge-backward node propagation described above, higher-order relation modeling is performed to obtain the initial values ​​of higher-order node representations under the auxiliary graph view. ; Step S40534: Perform graph convolution propagation based on the initial value of the higher-order node representation: Use the initial value of the higher-order node representation obtained by the hypergraph global encoder as the node feature input of the corresponding GCN encoder, and combine it with the normalized sub-adjacency matrix of the corresponding view to perform neighborhood aggregation and structure propagation to obtain the higher-order structure representation under the original graph view and the auxiliary graph view. Specifically, for the original graph view, higher-order nodes represent initial values. Adjacency matrix of the original graph Input the GCN encoder in the original graph view encoding branch to obtain the higher-order structure representation under the original graph view: ; in, The Behavior , represents the higher-order structure embedding of node i obtained after modeling the higher-order relations of the hypergraph and propagating the GCN structure under the original graph view.

[0085] Similarly, for auxiliary graph views, higher-order nodes are represented as initial values. Adjacency matrix of auxiliary graph Input the GCN encoder in the auxiliary graph view encoding branch to obtain the higher-order structure representation under the auxiliary graph view: ; in, The Behavior , representing a node The high-order structure embedding obtained after hypergraph high-order relation modeling and GCN structure propagation under the auxiliary graph view; Therefore, for any node It can simultaneously obtain node local embeddings Embedding of high-order structure in original diagram view And the embedding of higher-order structures in auxiliary diagrams This allows for the joint representation construction of node local information and dual-view high-order structural information.

[0086] Step S4054: Based on the obtained node local embedding representation of each node Original diagram view structure embedding and the embedding of auxiliary diagram view structures Perform supervised discrimination and calculate the connection probability estimate of node pairs; combine the node pair's true label to calculate the supervised discrimination loss. For batch Any node pair Extract the local representations of the two endpoint nodes respectively. and Based on the dimension-wise absolute difference between the two, node pair discrimination features are constructed, and the connection probability estimates of the node pairs are output. Its form can be expressed as: ; in, and To determine the parameters, For the Sigmoid function, Represents node pairs The predicted connection probability.

[0087] Based on node pairs of real labels That is, the actual connection relationships between nodes. The supervised discriminant loss is defined as: ; Supervised discriminant loss is used to directly constrain the model's ability to distinguish relationships between node pairs.

[0088] Step S4055: Perform collaborative constraints on the local node embedding with the original graph view structure embedding and the auxiliary graph view structure embedding, respectively. At the same time, perform cross-view collaborative constraints on the original graph view structure embedding and the auxiliary graph view structure embedding, and construct and calculate the local-global collaborative loss.

[0089] Specifically, let the set of deduplication nodes corresponding to the current batch be . For any node Local embedding of nodes is obtained respectively. Original diagram view structure embedding Embedding auxiliary diagram view structure To eliminate the influence of different embedding vector scales, the above embeddings are normalized: ; in, To prevent stable terms with a denominator of zero.

[0090] The similarity between two normalized embeddings is defined as ,in, The local embedding of the same node is considered to have a positive correspondence with the original graph view structure embedding, while the local embedding is considered to have a negative correspondence with the original graph view structure embedding of other nodes in the batch. The local-original graph collaborative loss is defined as: ; in, The temperature coefficient is used. This loss makes the local representation of the same node close to its higher-order structural representation formed under the true topology of the original graph, while maintaining the distinguishability of corresponding representations between different nodes.

[0091] The local embedding of the same node and the auxiliary graph view structure embedding are considered to have a positive correspondence, while the local embedding and the auxiliary graph view structure embedding of other nodes in the batch are considered to have a negative correspondence. The local-auxiliary graph collaborative loss is defined as: ; This loss ensures that the local representation of the same node is close to its higher-order structural representation formed under the latent topology of the auxiliary graph, thus maintaining consistency between the latent structural information supplemented by the auxiliary graph and the semantics of the node itself. The embedding of the same node in both the original and auxiliary graph views is considered a positive cross-view correspondence, while the cross-view embedding between different nodes is considered a negative correspondence. The cross-view collaborative loss is defined as: ; This loss ensures that the original graph view and the auxiliary graph view form a consistent basic semantic representation for the same node, while maintaining the distinguishability between cross-view representations of different nodes. The denominators of each of the above losses all include positive correspondence terms for the same node and negative correspondence terms for other nodes. Corresponding perfectly. Corresponding to negative correspondences. Ultimately, the local-global collaborative loss is expressed as: ; in, , and These are the weighting coefficients for the three collaborative loss terms. When the three loss terms have the same weight, the three weighting coefficients can be set to 1. .

[0092] Step S4056: Based on the node-to-hyperedge association matrix in the original graph view. Construct structural consistency labels for node pairs; calculate the similarity between the local embeddings of two nodes in a node pair; and construct and calculate the structural consistency loss based on the structural consistency labels and the similarity between the local embeddings. Specifically, based on the node-hyperedge association matrix in the original graph view. Construct structural consistency constraints. Let the incidence matrix from nodes to hyperedges be defined. The association vectors corresponding to node i and node j are respectively and Then, the structural label for whether a pair of order nodes shares at least one hyperedge is defined as: ; in, The label representing the structural consistency of node pair (i,j) and These represent the association strength between node i and node j and the e-th hyperedge in the original graph view, respectively.

[0093] Meanwhile, the cosine similarity between node pairs of local embeddings is: ; Further construct the structural consistency loss: ; in, is the similarity scaling factor. This structural consistency loss makes nodes sharing hyperedges more similar in the local representation space, thereby improving the consistency between the local representation and the higher-order structure of the original graph; Step S4057: Calculate the total loss of the first stage based on the supervised discriminant loss, the local-global collaborative loss and the structural consistency loss. Iteratively update the original graph-view encoding branch, the auxiliary graph-view encoding branch and the supervised discriminant parameters based on the total loss of the first stage to obtain the dual-view representation learning model trained in the first stage. Specifically, based on supervised discrimination loss, collaborative representation loss, and structural consistency loss, the total loss function for the first stage is defined as follows: ; in, , and These represent the weight coefficients of the supervision discrimination term, the collaborative representation term, and the structural consistency term, respectively.

[0094] During the training process, through Iterative optimization is performed to complete the joint modeling of local representation, original graph high-order representation, and auxiliary graph high-order representation. The local encoder, original graph view encoding branch, auxiliary graph view encoding branch, and supervised discriminant parameters are updated. After the first stage of training is completed, the updated dual-view representation is used to learn the model parameters for the next step to perform cross-view alignment, shared information learning, and private information decoupling across the entire graph.

[0095] Step S406, Second Phase Training: As Figure 2 As shown, the node position representation, the normalized adjacency matrix of the original graph, and the normalized adjacency matrix of the auxiliary graph are input into the dual-view representation learning model trained in the first stage across the entire graph to obtain the local embedding representation of the nodes in the entire graph, the structural embedding representation of the original graph view, and the structural embedding representation of the auxiliary graph view. Cross-view alignment, shared representation learning, and private representation decoupling are performed on the representation of the same node in the original graph view and the auxiliary graph view to ensure that the same node maintains semantic consistency in different views and to separate the unique structural information of different views, thus obtaining the dual-view representation learning model trained in the second stage. In the second phase of training, instead of processing node-to-batch batches, the entire graph node set is used directly. Given the input range, normalize the adjacency matrix using the original graph. Auxiliary graph normalized adjacency matrix Node position representation matrix and optional node attribute matrix The second stage of training is conducted. The purpose of the second stage of training is to further constrain the overall correspondence between the original and auxiliary views from a holistic perspective, ensuring consistency between them at the level of shared information while maintaining decoupling at the level of private information, thereby obtaining a dual-view representation with a clearer structure and stronger discriminative power.

[0096] The second phase of training performs cross-view alignment, shared information learning, and private information decoupling between the original graph view and the auxiliary graph view across the entire graph. The second phase of training includes the following steps: Step S4061: Input the node position representation and the normalized adjacency matrix of the original graph and the auxiliary graph into the dual-view representation learning model, and perform local encoding and dual-view high-order structure encoding respectively to obtain the local embedding representation, the original graph view structure embedding representation, the auxiliary graph view structure embedding representation and the corresponding node-hyperedge association matrix. The node position representations, the normalized adjacency matrix of the original graph, and the normalized adjacency matrix of the auxiliary graph are input into the dual-view representation learning model trained in the first stage across the entire graph; the local embedding representation matrix is ​​extracted through a local encoder. Higher-order structure encoding is performed through both the original graph-view encoding branch and the auxiliary graph-view encoding branch. Each view encoding branch first obtains initial values ​​for higher-order node representations and node-hyperedge association matrices via a hypergraph global encoder. Then, combined with the normalized adjacency matrix of the corresponding view, it undergoes graph convolution propagation via a GCN encoder to obtain the original graph-view structure embedding representation matrix. Auxiliary diagram view structure embedding representation matrix And the node-hyperedge incidence matrices of the original graph and the auxiliary graph. and ; ; ; ; in, , , These represent the local embedded representation, the original view representation, and the auxiliary view representation within the entire map area, respectively.

[0097] Step S4062: Based on the node-hyperedge association matrix, perform weighted aggregation on the node representations associated with the same hyperedge to construct local representations and higher-order representations corresponding to the original graph view and auxiliary graph view respectively; Specifically, to enhance the consistency between local representations and group-level higher-order structures, we first base our approach on the node-hyperedge correlation matrix. Construction of representation: For any hyperedge e, the local representations of the nodes associated with it under the original graph view are aggregated according to the association weights in the node-hyperedge association matrix to obtain the local representation. : ; in, This represents the strength of the association between node i and hyperedge e in the original graph view. To prevent stable terms with a denominator of zero; High-order representation of the original graph in the original graph view and auxiliary graph view Auxiliary graph higher-order representation They are defined as follows: , ; in, This indicates the strength of the association between node i and hyperedge e in the auxiliary graph view. and Let represent the group-level higher-order representations of the e-th hyperedge in the original graph view and the auxiliary graph view, respectively; The feature dimension representing the embedding of higher-order structures; This is a numerically stable term used to prevent the denominator from being zero or close to zero when the total association weight between a superedge and the nodes in the current batch is too small; superscript and These represent the original view and the auxiliary view, respectively.

[0098] Step S4063: Treat the local representation under the same hyperedge as a positive correspondence with the higher-order representation of the original graph or the higher-order representation of the auxiliary graph, and treat the local representations of different hyperedges as a negative correspondence with the representations of the original graph or the auxiliary graph, and construct and calculate the contrast loss. Specifically, respectively and , and We treat the representations between different hyperedges as positive corresponding samples and construct a contrastive loss.

[0099] Let the temperature coefficient be The contrast loss under the original image view It can be written as: ; in, This indicates the preset number of hyperedges, that is, the total number of potential hyperedges set in the hypergraph global encoder. Contrast loss under the auxiliary graph view. Similarly, we can define and obtain the consistent comparison loss: ; in, This represents the similarity function between normalized vectors.

[0100] The consistency loss in this embodiment ensures that the local representation remains consistent with the higher-order structure of the two-view at the hyperedge level, thereby enhancing the stability of the group-level topological semantics.

[0101] This embodiment addresses the problem that existing graph representation learning methods mainly rely on node-edge relationships or local adjacency propagation, which struggles to fully characterize higher-order structural relationships such as group interactions, shared memberships, and cross-regional dependencies. It establishes a hypergraph global encoder in both the original graph view encoding branch and the auxiliary graph view encoding branch. Based on the association between nodes and hyperedges, multiple semantically or structurally related nodes are organized into a hyperedge set. Multi-node joint relationships are modeled through aggregation propagation from nodes to hyperedges and from hyperedges to nodes. Therefore, the model is no longer limited to explicit connections between pairs of nodes but can extract higher-order dependency information between nodes at the group-level structure. This enhances the ability to express complex topological relationships, potential shared membership structures, and cross-regional associations, improving the global structural awareness of node representations and the discrimination performance in downstream tasks.

[0102] Step S4064, Shared-private representation encoding and cross-view shared alignment: Based on the shared encoder and private encoder, the shared representation and private representation of the original graph view and the auxiliary graph view are extracted, and the cross-view shared information alignment loss is constructed through cross-view contrastive learning; Based on the original graph view structure embedding representation and the auxiliary graph view structure embedding representation obtained from the aforementioned steps, the structure embedding representations of the two views are respectively input into a shared encoder with shared parameters and a private encoder that are independent of each other, to obtain the shared representation and private representation corresponding to the original graph view and the auxiliary graph view. Furthermore, positive correspondence pairs are constructed using the shared representations of the same node in both views, and negative correspondence pairs are constructed using the shared representations of different nodes in both views. This constructs a cross-view shared information alignment loss, ensuring that the shared information corresponding to the same node in both views remains consistent.

[0103] Specifically, let the structural embedding representation matrices output by the aforementioned original graph-view encoding branch and auxiliary graph-view encoding branch be: ; in, The matrix represents the higher-order structure of the original graph view. The auxiliary diagram view represents a higher-order structure matrix. Indicates the number of nodes.

[0104] Introducing a shared encoder for parameter sharing between two views. And private encoders with independent parameters, corresponding to the original view and the auxiliary view respectively. and Encode them separately: ; ; in, and These represent the shared representation matrices of the original view and the auxiliary view, respectively. and Let the private representation matrices of the original view and the auxiliary view be respectively, and let the (i)th row of the above four representation matrices be respectively ; Shared encoder Use the same network parameters between the two views to extract common information from both views; private encoder and Independent network parameters are used to extract view-specific information from the original graph view and the auxiliary graph view separately. To ensure consistency in the shared representation of the same node in both views, [the following parameters are used]. As a positive correspondence pair, As a negative correspondence, it represents a pair, where Construct a shared alignment loss from the original graph view to the auxiliary graph view. : ; in, Represents the similarity function. The shared temperature coefficient is used to represent the alignment process.

[0105] Similarly, construct a shared alignment loss for pointing from the auxiliary graph view to the original graph view: ; Therefore, the cross-view shared information alignment loss is defined as: By minimizing alignment loss through cross-view shared information This allows the shared representations of the same node in the original graph view and the auxiliary graph view to be close to each other, while maintaining the distinguishability of the shared representations of different nodes across views, thereby enhancing the consistency of common information between the two views.

[0106] Step S4065: Apply orthogonal constraints to the shared and private representations under the same view, and apply separation constraints to the private representations of different views, and construct the shared-private orthogonal decoupling loss and the cross-view private information separation loss respectively; This step applies orthogonal decoupling constraints to the shared and private representations in the original and auxiliary views, which can reduce the redundant correlation between shared and private information within the same view. At the same time, applying cross-view separation constraints to the private representations in the original and auxiliary views can reduce the repeated encoding of the same information by the private representations of the two views and enhance the ability of the private representations to characterize the special structural information of the corresponding views. Specifically, first, row normalization is performed on the shared representation matrix and the private representation matrix respectively, resulting in: ; For any representation matrix The normalized representation is: ; in, This is a numerically stable term used to avoid the denominator being zero during normalization.

[0107] For the original graph view, by constraining the cross-correlation between shared and private representations, a shared-private orthogonal decoupling loss is defined under the original graph view. for: ; Similarly, the shared-private orthogonal decoupling loss under the auxiliary graph view is defined as: ; Therefore, the shared-private orthogonal decoupling loss is defined as: ; By minimizing, the correlation between shared and private representations in the same view is reduced, so that the shared representation mainly retains the common information in the two views, and the private representation mainly retains the specific information in the corresponding view.

[0108] Furthermore, to reduce the redundant encoding of the same information by the private representations of the original graph view and the auxiliary graph view, a cross-view separation constraint is imposed on the private representations in the two views, and the cross-view private information separation loss is defined as: ; By minimizing This reduces the correlation between the private representation of the original graph and the private representation of the auxiliary graph, making the private representation of the original graph focus more on characterizing the observed topology and its unique structural information in the original graph, and making the private representation of the auxiliary graph focus more on characterizing the potential structural information supplemented by the auxiliary graph, thereby enhancing the difference and complementarity between the two private representations.

[0109] Furthermore, based on the shared-private orthogonal decoupling loss and the cross-view private information separation loss, a private information decoupling loss is constructed: ; in, and The weight coefficients represent the shared-private orthogonal decoupling loss and the cross-view private information separation loss, respectively. On the one hand, this reduces the information redundancy between shared and private representations within the same view, and on the other hand, it reduces the duplicate information between private representations in different views, so that shared and private representations can respectively undertake the functions of cross-view public information modeling and view-specific information modeling.

[0110] After obtaining the shared and private representations of each view, the shared and private representations under the same view are combined to obtain the original graph-view combined representation and the auxiliary graph-view combined representation: ; Where | represents the feature concatenation operation. The original diagram view combination represents a matrix. The auxiliary graph-view combination representation matrix is ​​used. Both the original graph-view combination representation and the auxiliary graph-view combination representation retain shared information between the two views, as well as the unique structural information corresponding to each view, and serve as the representation input for subsequent fusion and target task prediction.

[0111] In the subsequent second-stage overall objective function, the cross-view shared information alignment loss is considered. Decoupling loss from private information Each loss term is treated as an independent loss term and participates in joint optimization to maintain a one-to-one correspondence between the loss term names, mathematical symbols, and their technical effects.

[0112] Step S4066: Compare the loss Cross-view information sharing alignment loss Cross-view orthogonal decoupling loss and loss of private information decoupling The weighted combination forms the total loss for the second stage of training. Based on the total loss for the second stage of training, the parameters of the original graph-view encoding branch and the auxiliary graph-view encoding branch are iteratively updated to obtain the dual-view representation learning model after the second stage of training. After pre-training, the shared embedding representations and private representations in the original graph view and the auxiliary graph view are combined to obtain: ; in, This indicates a feature splicing operation.

[0113] In the second phase of training, the combined original graph is represented. and auxiliary diagram view representation The input is then fed into the subsequent fusion and decoding module to obtain the prediction results for the target task.

[0114] The overall objective function for the second stage of training is defined as: ; in, , These represent the weighting coefficients.

[0115] Through optimization The second-stage model can fuse the original graph-view representation with the auxiliary graph-view representation while preserving cross-view shared information and view-specific information, and complete the training of the corresponding downstream tasks. Through the trained dual-view representation learning model, the embedding matrix of the original graph-view structure can be obtained. Auxiliary diagram view structure embedding matrix Original graph view shared representation matrix Auxiliary diagram view shared representation matrix Original graph view private representation matrix Auxiliary graph view private representation matrix And the original diagram representation matrix after combination. And auxiliary diagram view representation matrix .

[0116] This embodiment addresses the problems of ineffective differentiation between shared and private information and semantic aliasing in cross-view representation learning in existing dual-view representations. In the second stage of training, this embodiment uses all-graph nodes as objects to perform cross-view alignment on the original graph view representation and the auxiliary graph view representation. Shared projection maps common semantics from different views to a unified shared subspace, ensuring consistency in the shared representation of the same node across different views. Simultaneously, the dual-view representation is further decomposed into shared and private representations. Shared representation alignment constraints strengthen cross-view common information, cross-view private representation orthogonality constraints reduce redundant representations of private information between the two views, internal view decoupling constraints reduce redundant coupling between shared and private components within the same view, and variance regularization constraints prevent representation collapse. Therefore, information aliasing between the original graph view and the auxiliary graph view is reduced, the consistency of shared semantic expression and the independence of private structural information are improved, resulting in a clearer structural hierarchy and stronger discriminative ability in the final node representation.

[0117] After completing the two-stage training, the model is switched to inference mode, and the entire graph node set is used. For objects, represent matrices based on node positions. Optional node attribute matrix Normalized adjacency matrix of the original graph and the normalized adjacency matrix of the auxiliary graph Perform forward encoding on all nodes once, and output local embedding, original graph view structure embedding, and auxiliary graph view structure embedding.

[0118] In step 4, the node position representations in the original graph view and the auxiliary graph view, the normalized adjacency matrix of the original graph, and the normalized adjacency matrix of the auxiliary graph are input into the trained dual-view representation learning model for dual-view node encoding, resulting in node local embeddings, original graph view structure embeddings, and auxiliary graph view structure embeddings. The input data also includes an optional node attribute matrix, comprising the following steps: Step 41: Input the node position representation matrix and the optional node attribute matrix into the local encoder of the trained dual-view representation learning model. The local encoder encodes the node's own features and its local structural information to obtain the node local embedding matrix. ; Step 42: Input the node position representation matrix, optional node attribute matrix, and the normalized adjacency matrix of the original graph into the original graph view encoding branch. Model the higher-order relationships of nodes and hyperedges through the hypergraph global encoder to obtain the initial values ​​of the higher-order node representations. Then, the GCN encoder combines the normalized adjacency matrix of the original view to perform graph convolution propagation to obtain the original graph view structure embedding matrix. ; Step 43: Input the node position representation matrix, optional node attribute matrix, and auxiliary graph normalized adjacency matrix into the auxiliary graph view encoding branch of the trained dual-view representation learning model. Model higher-order relationships between nodes and hyperedges using the hypergraph global encoder to obtain initial values ​​for higher-order node representations. Then, perform graph convolution propagation using the GCN encoder in conjunction with the normalized adjacency matrix of the auxiliary view to obtain the original graph view structure embedding matrix. ; For any node The local embeddings of their nodes are obtained respectively. Original diagram view structure embedding Embedding auxiliary diagram view structure Stacking the above three types of embeddings of all nodes sequentially according to their node numbers, we get: ; ; ; in, Indicates the total number of nodes. Indicates the node embedding dimension; This represents the local embedding matrix of a node, used to characterize the node's own positional semantics and optional attribute information; The original graph-view structure embedding matrix is ​​used to characterize the known topological relationships and higher-order structural information in the original graph-view. The auxiliary graph view structure embedding matrix is ​​used to characterize the potential relationships and higher-order structural information supplemented by the auxiliary graph view.

[0119] All three types of embeddings correspond to the same set of nodes. And the first of each matrix The rows correspond to the nodes respectively Local embedding, original graph-view structure embedding, and auxiliary graph-view structure embedding; Through the aforementioned dual-view node encoding process, the original graph view encoding branch first uses a hypergraph global encoder to model the higher-order node relationships in the original graph, and then uses a GCN encoder combined with the normalized adjacency matrix of the original graph for structure propagation to obtain the original graph view structure embedding. The auxiliary graph view encoding branch first uses a hypergraph global encoder to model the potential higher-order relationships in the auxiliary graph, and then uses a GCN encoder combined with the normalized adjacency matrix of the auxiliary graph for structure propagation to obtain the auxiliary graph view structure embedding. Thus, the model can simultaneously obtain the local semantic information of the nodes themselves, the known topological and historical association information in the original graph, and the potential structural association information supplemented by the auxiliary graph. In one implementation, the local node embedding, the original graph view structure embedding, and the auxiliary graph view structure embedding of the same node are concatenated along the feature dimension to obtain the first... The final representation of each node: ; in, This indicates that a vector concatenation operation is performed along the feature dimension. Represents a node The final node representation. Accordingly, the final representations of all nodes are stacked sequentially according to their node numbers to obtain the final node representation matrix: ; For the task of identifying potential fault associations in industrial equipment, candidate node pairs are input into a pre-defined association discrimination module. The module calculates the association probabilities between equipment components and fault types, operating parameters and fault types, sensors and fault types, or different equipment components. Based on the comparison between the association probabilities and a pre-defined discrimination threshold, it determines whether the corresponding candidate node pairs constitute a potential fault association. The identified potential fault associations are then used for fault diagnosis, equipment inspection, or maintenance decisions.

[0120] Example 2 Based on Embodiment 1, this embodiment provides a graph structure representation learning system based on conditional graph and dual-view decoupling, including: The original graph view construction module is configured to acquire the graph data to be processed in the task scenario and construct an original graph view including a set of nodes and a set of edges. The structure condition generation module is configured to extract the node position representation of the nodes in the original graph view, and construct the structure condition vector of each node based on the node degree, average neighbor degree and graph density statistics based on the original graph view, thus forming the structure condition matrix. The auxiliary graph view generation module is configured to input the node position representation and the structural condition matrix into a trained conditional diffusion network to generate an auxiliary node feature matrix; and to select candidate neighbors for each node based on the similarity between the auxiliary node features to construct the auxiliary graph view. The dual-view node encoding module is configured to input the node position representations in the original graph view and the auxiliary graph view, the normalized adjacency matrix of the original graph and the normalized adjacency matrix of the auxiliary graph into the trained dual-view representation learning model to encode the dual-view nodes, thereby obtaining the node local embedding, the original graph view structure embedding and the auxiliary graph view structure embedding. The fusion module is configured to fuse the local node embedding, the original graph view structure embedding, and the auxiliary graph view structure embedding to obtain the final node representation, and to execute the downstream tasks of the task scenario based on the final node representation.

[0121] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0122] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0123] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A graph structure representation learning method based on decoupling of conditional graphs and dual views, characterized in that, Includes the following steps: Obtain the graph data to be processed in the task scenario and construct the original graph view including the node set and edge set; Extract the node position representation of the nodes in the original graph view, and construct the structure condition vector of each node based on the node degree, average neighbor degree and graph density statistics based on the original graph view, thus forming the structure condition matrix; The node position representation and the structural condition matrix are input into the trained conditional diffusion network to generate the auxiliary node feature matrix. Candidate neighbors are selected for each node based on the similarity between the features of auxiliary nodes, and an auxiliary graph view is constructed. The node position representations in the original graph view and the auxiliary graph view, the normalized adjacency matrix of the original graph and the normalized adjacency matrix of the auxiliary graph are input into the trained dual-view representation learning model to encode the dual-view nodes, resulting in node local embedding, original graph view structure embedding and auxiliary graph view structure embedding. The node local embedding, the original graph view structure embedding, and the auxiliary graph view structure embedding are fused to obtain the final node representation, and the downstream tasks of the task scenario are executed based on the final node representation.

2. The graph structure representation learning method based on conditional graph and dual-view decoupling as described in claim 1, characterized in that, The node degree is the number of connected nodes of node i in the original graph view; The average degree of node i's neighbors is the average degree of all neighboring nodes in the neighbor set of node i. graph density of the original graph view It is the ratio between the actual number of edges in the original graph view and the maximum number of edges that can theoretically exist in the original graph view.

3. The graph structure representation learning method based on conditional graph and dual-view decoupling as described in claim 1, characterized in that, The node position representation and the structural condition matrix are input into the trained conditional diffusion network to generate an auxiliary node feature matrix. Based on the similarity between the auxiliary node features, candidate neighbors are selected for each node to construct an auxiliary graph view, including the following steps: For each node, a random noise vector is sampled from a standard Gaussian distribution as the noise state of the node at the initial diffusion step T of the reverse generation process; Based on the noise state at the initial diffusion step T, according to the sampling step arrive The sequential execution conditional backsampling, for the first One diffusion step, the current noise state diffusion step and the structural condition vector of node i Input the trained conditional diffusion network to obtain the estimated state of node i in the previous diffusion step. ; Based on the estimated state of the previous diffusion step, the node state is iteratively updated according to the conditional backsampling rule to obtain the auxiliary node features; Calculate the similarity between the auxiliary node features of any two nodes, and then sort them according to the similarity. Sort the nodes from highest to lowest similarity and select the top K nodes with the highest similarity to form the candidate neighbor set for node i. Construct an auxiliary adjacency matrix based on the candidate neighbor set. ; Determining the set of auxiliary edges based on the auxiliary adjacency matrix. Constructing auxiliary graph views .

4. The graph structure representation learning method based on conditional graph and dual-view decoupling as described in claim 1, characterized in that, The training process of a conditional diffusion network includes the following steps: Obtaining the node position representation matrix: based on the training graph The nodes in the training graph are encoded to obtain a node position representation matrix; For training images For any node i in the array, calculate its degree. And determine the neighbor set of node i; calculate the average neighbor degree of node i based on the node degree of each neighbor node in the neighbor set of node i. Calculate the graph density δ based on the number of nodes and edges in the training graph; construct the structure condition vector of node i based on node degree, average neighbor degree, and graph density. Set diffusion step size and noise scheduling parameters And construct forward-noise samples; During the reverse generation process, the construction parameters are: Conditional denoising network The original position features of nodes are recovered from the forward-noiseed samples by using a conditional denoising network. The conditional denoising network is trained with the goal of minimizing the mean square error between the estimated value and the true position representation, and the trained conditional denoising network is obtained.

5. The graph structure representation learning method based on conditional graph and dual-view decoupling as described in claim 1, characterized in that, The dual-view representation learning model includes a parallel local encoder, an original graph-view encoding branch and an auxiliary graph-view encoding branch, and a projection fusion head; The local coding module is used to extract the semantic information and state features of the node itself; The original view coding branch and the auxiliary view coding branch are used to model the structural relationships in the original view and the auxiliary view, respectively; Both the original graph view coding branch and the auxiliary graph view coding branch adopt a concatenated coding structure, including a hypergraph global encoder and a GCN encoder; The Hypergraph Global Encoder is used to model higher-order relationships based on node-hyperedge relationships in the view, and to obtain initial values ​​of higher-order node representations containing higher-order semantic relationships. The GCN encoder takes the initial value of the higher-order node representation as input and combines it with the normalized adjacency matrix of the corresponding view to perform structure propagation, thereby obtaining the structure embedding of the corresponding view. The projection fusion head is used to fuse the local node embedding, the original graph view structure embedding, and the auxiliary graph view structure embedding to obtain the final node representation.

6. The graph structure representation learning method based on conditional graph and dual-view decoupling as described in claim 5, characterized in that, The dual-view representation learning model also includes a sample construction and sampling module and a two-stage training unit; The sample construction and sampling module is used to construct positive sample node pairs based on the random walk mechanism of the original graph, and to mine difficult negative samples based on the auxiliary graph structure. At the same time, it combines a random negative sampling strategy to form a complete training sample set; and further generates a subgraph structure input for mini-batch training. The two-stage training unit is configured to optimize the dual-view representation learning model using a two-stage training strategy.

7. The graph structure representation learning method based on conditional graph and dual-view decoupling as described in claim 5, characterized in that, The steps for training a dual-view representation learning model include the following: The acquired training data is partitioned and a training graph is constructed using the original graph view. For the original graph view in the training graph, an auxiliary graph view is constructed: the node position representation of the nodes in the original graph view is extracted, and the structure condition vector of each node is constructed based on the node degree, average neighbor degree and graph density statistics based on the original graph view, forming a structure condition matrix; the node position representation and the structure condition matrix are input into the trained conditional diffusion network to generate the auxiliary node feature matrix. Candidate neighbors are selected for each node based on the similarity between the features of auxiliary nodes, and an auxiliary graph view is constructed. Positive sample node pairs are constructed by performing a random walk mechanism based on the original graph view, and difficult negative sample node pairs are mined based on the auxiliary graph structure. At the same time, a node pair training set is formed by combining a random negative sampling strategy. The adjacency matrices of the original graph view and the auxiliary graph view are normalized respectively to obtain the normalized adjacency matrices. and , and the batch subgraph structure corresponding to the training batch; The node connection relationship is learned by the supervised discriminative loss constraint model of positive and negative sample node pairs, and the first stage of training of the dual-view representation learning model is carried out by the collaborative constraint between the local representation and the dual-view higher-order representation. In the first stage of training, the node position representation, the normalized adjacency matrix of the original graph, and the normalized adjacency matrix of the auxiliary graph are input into the dual-view representation learning model across the entire graph to obtain the local embedding representation, the original graph view representation, and the auxiliary graph view representation of the nodes across the entire graph. Cross-view alignment, shared representation learning, and private representation decoupling are performed on the representations of the same node in the original graph view and the auxiliary graph view to ensure that the same node maintains semantic consistency in different views and to separate the unique structural information of different views, thus obtaining the dual-view representation learning model after the second stage of training.

8. The graph structure representation learning method based on conditional graph and dual-view decoupling as described in claim 7, characterized in that, The first phase of training includes the following steps: Based on the batch subgraph structure, in each round of mini-batch training, a batch of node pairs is selected from the node pair training set, and the node position representation submatrix corresponding to the current batch is extracted. Original graph subadjacency matrix and auxiliary graph subadjacency matrix ; Submatrix based on node position representation The input is fed into the local encoder of the corresponding view encoding branch to perform fine-grained feature encoding on the node's own features and its local neighborhood structure, thus obtaining the node's local embedding. For the original graph view and the auxiliary graph view, the hypergraph encoder and GCN encoder of the corresponding branches are used to perform hypergraph high-order relation modeling and GCN structure propagation to obtain the original graph view structure embedding and the auxiliary graph view structure embedding. Supervised discrimination is performed based on the obtained local embedding representation of each node, the original graph, and the auxiliary graph view structure embedding, and the connection probability estimate of node pairs is calculated. The supervised discrimination loss is calculated by combining the nodes with the real labels; The local node embedding is coordinated with the original graph view structure embedding and the auxiliary graph view structure embedding, respectively. At the same time, cross-view coordination constraints are applied to the original graph view structure embedding and the auxiliary graph view structure embedding. The local-global coordination loss is constructed and calculated. Based on the node-to-hyperedge association matrix in the original graph view Construct structural consistency labels for node pairs; calculate the similarity between the local embeddings of two nodes in a node pair; and construct and calculate the structural consistency loss based on the structural consistency labels and the similarity between the local embeddings. The first-stage total loss is calculated based on the supervised discriminant loss, the local-global collaborative loss, and the structural consistency loss. The original graph-view encoding branch, the auxiliary graph-view encoding branch, and the supervised discriminant parameters are iteratively updated based on the first-stage total loss to obtain the dual-view representation learning model trained in the first stage.

9. The graph structure representation learning method based on conditional graph and dual-view decoupling as described in claim 7, characterized in that, The second phase of training includes the following steps: The node position representation and the normalized adjacency matrix of the original graph and the auxiliary graph are input into the dual-view representation learning model. Local encoding and dual-view higher-order structure encoding are performed respectively to obtain the local embedding representation, the original graph view structure embedding representation, the auxiliary graph view structure embedding representation and the corresponding node-hyperedge association matrix. Based on the node-hyperedge association matrix, the node representations associated with the same hyperedge are weighted and aggregated to construct local representations and higher-order representations corresponding to the original graph view and auxiliary graph view, respectively. The local representation under the same hyperedge is considered to have a positive correspondence with the higher-order representation of the original graph or the higher-order representation of the auxiliary graph, while the local representations of different hyperedges are considered to have a negative correspondence with the representations of the original graph or the auxiliary graph. The contrast loss is then constructed and calculated. Based on the shared encoder and the private encoder, the shared representation and the private representation of the original graph view and the auxiliary graph view are extracted, and the cross-view shared information alignment loss is constructed through cross-view contrastive learning. Orthogonal constraints are applied to shared and private representations under the same view, and separation constraints are applied to private representations in different views. Shared-private orthogonal decoupling loss and cross-view private information separation loss are constructed respectively. The contrast loss, cross-view shared information alignment loss, cross-view orthogonal decoupling loss, and private information decoupling loss are weighted and combined to form the total training loss for the second stage. Based on the total training loss for the second stage, the parameters of the original graph-view encoding branch and the auxiliary graph-view encoding branch are iteratively updated to obtain the dual-view representation learning model after the second stage training.

10. A graph structure representation learning system based on conditional graphs and dual-view decoupling, characterized in that, include: The original graph view construction module is configured to acquire the graph data to be processed in the task scenario and construct an original graph view including a set of nodes and a set of edges. The structure condition generation module is configured to extract the node position representation of the nodes in the original graph view, and construct the structure condition vector of each node based on the node degree, average neighbor degree and graph density statistics based on the original graph view, thus forming the structure condition matrix. The auxiliary graph view generation module is configured to input the node position representation and the structural condition matrix into the trained conditional diffusion network to generate an auxiliary node feature matrix. Candidate neighbors are selected for each node based on the similarity between the features of auxiliary nodes, and an auxiliary graph view is constructed. The dual-view node encoding module is configured to input the node position representations in the original graph view and the auxiliary graph view, the normalized adjacency matrix of the original graph and the normalized adjacency matrix of the auxiliary graph into the trained dual-view representation learning model to encode the dual-view nodes, thereby obtaining the node local embedding, the original graph view structure embedding and the auxiliary graph view structure embedding. The fusion module is configured to fuse the local node embedding, the original graph view structure embedding, and the auxiliary graph view structure embedding to obtain the final node representation, and to execute the downstream tasks of the task scenario based on the final node representation.

Images