High-order structure sensing hypergraph convolutional network construction method and device, medium and equipment

Abstract

The invention relates to the technical field of graph neural networks, in particular to a high-order structure perception hypergraph convolutional network construction method and device, a medium and equipment, and the method comprises the steps: generating candidate sub-graphs through frequent sub-graph mining based on an original graph structure, introducing a sequential embedding mechanism to perform structure representation learning and screening on the candidate sub-graphs to obtain frequent sub-graphs, dynamically generating a frequent sub-graph set, constructing a high-order hyperedge set based on the frequent sub-graph set, and constructing a hyperedge structure based on the high-order hyperedge set; and performing feature aggregation and representation learning of high-order structure semantics on node features in the hypergraph structure by adopting a hypergraph convolutional network, and generating enhanced node representation. According to the method, the reasonability, representativeness and task correlation of high-order structure selection can be improved, then the expressivity, interpretability and robustness of the model are improved, and the method has wide application prospects and industrialization value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of graph neural networks, and particularly relates to a high-order structure perception hypergraph convolution network construction method, device, medium and equipment. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] With the rapid development of information technology and the explosive growth of data scale, graph neural networks as a kind of deep learning method capable of effectively modeling graph structure data have been widely applied in many fields such as social network analysis, bioinformatics, recommendation system, knowledge graph and traffic network. Graph neural networks (GNN) realize efficient representation learning of nodes or graphs as a whole by aggregating the features of neighboring nodes. However, the existing GNN model is based on simple graph construction, that is, only the pairwise relationship between nodes is considered, and it is difficult to fully express the more complex multi-element interaction relationship in the real world. In many practical applications, the connection between nodes often presents a group interaction mode, such as multi-author collaboration in scientific research cooperation, group chat in social platforms, joint purchase in e-commerce, etc., which exceeds the expression ability of simple graph structure.

[0004] In order to overcome the limitations of simple graph structure, hypergraph as a generalized graph model capable of describing high-order relationships between any number of nodes has gradually attracted attention. Hypergraph can naturally express high-order relationships such as group interaction and local structure mode by connecting multiple nodes through hyperedges. Therefore, data modeling based on hypergraph and research on hypergraph neural networks have become an important direction in the field of graph learning in recent years. Early methods such as HGNN define hypergraph convolution through spectral theory, realizing the aggregation of collaborative features of nodes within hyperedges; HyperGCN converts hypergraph into an approximate simple graph for convolution operation, improving the calculation efficiency; HGNN+, HyperSAGE, HNHN and other methods further introduce multi-scale feature extraction, attention mechanism, double-channel information fusion and other strategies to improve the expression ability of the model to complex structures. These methods have achieved good results in classification, clustering, ranking, anomaly detection and other tasks.

[0005] Despite the progress made by hypergraph neural networks in various task scenarios, the following problems still exist: First, most hypergraph neural network models rely on manually predefined high-order structural units (such as motifs and cliques). These structures heavily depend on expert knowledge or domain priors, making it difficult to adapt to the differences and diverse distributions of structural patterns across different datasets. Furthermore, manually defined high-order structural units lack flexibility and scalability, failing to meet the actual modeling needs of complex graph data and limiting the interpretability and generalization ability of the models to some extent. Second, current high-order structure construction often relies on static, rule-based generation strategies, such as nearest neighbor, clustering, or attribute consistency methods. These methods often fail to fully exploit the high-frequency high-order structural patterns implicit in the original graph data, resulting in insufficient modeling of high-order relationships. In addition, when facing large-scale, complex, and diverse graph data, the high-order structures constructed by existing methods are prone to information redundancy or insufficient coverage of local patterns, thus affecting the model's expressive power and downstream task performance.

[0006] Currently, academic and industrial communities both domestically and internationally have proposed various solutions to address the aforementioned problems. For example, some methods generate hyperedges based on distance or clustering strategies and use them for higher-order modeling of nearest neighbor relationships; others construct semantically consistent hyperedges using attribute labels or node features to enhance the semantic rationality of higher-order structures. However, most of these solutions are limited to utilizing shallow features or low-order structural information, failing to automatically mine complex higher-order patterns from data and exhibiting insufficient adaptability to complex, heterogeneous, or large-scale graph data. Some research has attempted to introduce data-driven mechanisms into graph neural networks, but systematic solutions are still lacking in areas such as higher-order structure recognition, pattern selection, and hyperedge construction. In particular, existing technologies still have significant shortcomings in how to automatically and efficiently extract high-frequency, representative higher-order subgraph patterns from raw graph data and use them to guide hypergraph structure construction.

[0007] In view of the above-mentioned defects, the present invention has made improvements. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, this invention provides a method, apparatus, medium, and device for constructing high-order structure-aware hypergraph convolutional networks. This method can improve the rationality, representativeness, and task relevance of high-order structure selection, thereby enhancing the expressiveness, interpretability, and robustness of the model. It has broad application prospects and industrialization value.

[0009] To achieve the above objectives, the present invention provides the following technical solution: Firstly, a method for constructing a high-order structure-aware hypergraph convolutional network is provided, the method comprising: Based on the original graph structure, candidate subgraphs are generated through frequent subgraph mining. A sequential embedding mechanism is introduced to learn the structural representation of the candidate subgraphs and filter them to obtain frequent subgraphs. A set of frequent subgraphs is dynamically generated. A set of higher-order hyperedges is constructed based on the set of frequent subgraphs, and a hypergraph structure is constructed based on the set of higher-order hyperedges. A hypergraph convolutional network is used to perform high-order structural semantic feature aggregation and representation learning on the node features in the hypergraph structure, and to generate enhanced node representations.

[0010] Furthermore, The process of generating candidate subgraphs based on the original graph structure through frequent subgraph mining includes: Based on the original graph structure, the neighborhood of a node is used as the basic unit. The neighborhood subgraphs within the local k-order neighborhood of a node are enumerated, and the candidate subgraph set is determined by statistically analyzing the frequency of occurrence of the neighborhood subgraphs in the whole graph.

[0011] Furthermore, The introduced sequential embedding mechanism performs structural representation learning and filtering on the candidate subgraphs to obtain frequent subgraphs, and dynamically generates a set of frequent subgraphs, including: Build node representations for each candidate subgraph and input them into a graph embedding model to learn their structural representations; The partial order distance function based on sequential embedding is used to determine the inclusion relationship between different subgraphs and whether different subgraphs constitute equivalence classes. Based on the judgment results, positive and negative sample pairs are constructed, and the graph embedding model is trained to make the embedding distance of similar subgraphs closer and the embedding distance of dissimilar subgraphs farther. Based on the embedding distance minimization criterion, the set of frequent subgraphs with the smallest error is dynamically selected from the candidate subgraph set.

[0012] Furthermore, The expression for the partial order distance function based on sequential embedding is:

[0013] in, The structure embedding function of the subgraph is represented. Characteristic subgraph The partial order relationship between them.

[0014] Furthermore, The training graph embedding model uses a discriminative ranking loss function for positive and negative sample pairs, the expression of which is:

[0015] in, Represents a pair of positive sample subgraphs that satisfy a partial order relation. For negative sample pairs that do not satisfy the partial order relation. This is the minimum interval constant between positive and negative sample pairs. Characteristic subgraph The partial order relation between them Characteristic subgraph and The partial order relationship between them.

[0016] Furthermore, The method of dynamically selecting the set of frequent subgraphs with the smallest error from the candidate subgraph set based on the embedding distance minimization criterion includes: Perform distance-compressed clustering on all embedded candidate subgraphs, and select the cluster centers representing the most embedded subgraphs based on the ordinal embedding distance. Each subgraph is a set of frequent subgraphs.

[0017] Furthermore, The method of dynamically selecting the set of frequent subgraphs with the smallest error from the candidate subgraph set based on the embedding distance minimization criterion also includes: An iterative greedy sampling strategy is adopted to gradually select the subgraph with the smallest error from the candidate subgraph set and add it to the frequent subgraph set.

[0018] Furthermore, The construction of a higher-order hyperedge set based on the frequent subgraph set includes: Based on the set of frequent subgraphs, for each selected frequent subgraph pattern, all matching instances of the pattern are enumerated in the original graph, and the set of nodes contained in each instance is used as a higher-order hyperedge to construct a set of higher-order hyperedges.

[0019] Furthermore, The step of employing a hypergraph convolutional network to perform high-order structural semantic feature aggregation and representation learning on the node features in the hypergraph structure, and generating enhanced node representations, includes: Initialize the node features and weights in the hypergraph structure; Perform feature aggregation from nodes to hyperedges and feature propagation from hyperedges back to nodes; By stacking multiple convolutional layers and using the output of each layer as the input of the next layer, an enhanced node representation is finally generated.

[0020] Secondly, a device for constructing a high-order structure-aware hypergraph convolutional network is also provided, the device comprising: The hypergraph construction module is used for the original graph structure. It generates candidate subgraphs through frequent subgraph mining, introduces a sequential embedding mechanism to learn the structural representation of the candidate subgraphs and filter them to obtain frequent subgraphs, dynamically generates a set of frequent subgraphs, constructs a set of higher-order hyperedges based on the set of frequent subgraphs, and constructs a hypergraph structure based on the set of higher-order hyperedges. The hypergraph convolution module is used to perform high-order structural semantic feature aggregation and representation learning on the node features in the hypergraph structure using a hypergraph convolutional network, and to generate enhanced node representations.

[0021] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing one or more programs, which, when executed, can realize the method for constructing a high-order structure-aware hypergraph convolutional network as described above.

[0022] Based on the same inventive concept, the present invention also provides an electronic device, including a processor, a communication interface, a computer-readable storage medium as described above, and a communication bus; wherein the processor, the communication interface, and the computer-readable storage medium communicate with each other through the communication bus; the processor is used to execute a program stored in the computer-readable storage medium.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention overcomes the limitations of existing hypergraph neural networks that rely excessively on artificial prior high-order structures and are difficult to adapt to complex heterogeneous graph data. Through data-driven frequent subgraph mining and sequential embedding filtering mechanisms, it realizes the automated construction of high-order structures and improves the model's adaptability and generalization ability to diverse graph data.

[0024] 2. The high-order hyperedge generation and hypergraph convolution co-design adopted in this invention effectively enhance the ability to aggregate high-order structural information of node features, and significantly improve the discriminativeness and robustness of node representation.

[0025] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings.

[0026] The invention will now be further described with reference to the accompanying drawings. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating a method for constructing a high-order structure-aware hypergraph convolutional network according to an embodiment of the present invention. Figure 2This is a schematic diagram of the overall structure of a high-order structure-aware hypergraph convolutional network based on frequent subgraph mining, according to an embodiment of the present invention. Figure 3 This is a schematic diagram of a sequential embedding process according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a high-order structure-aware hypergraph convolutional network construction device according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0030] like Figure 1 As shown, an embodiment of the present invention provides a method for constructing a high-order structure-aware hypergraph convolutional network, the method comprising the following steps: S1. Based on the original graph structure, candidate subgraphs are generated through frequent subgraph mining. After the candidate subgraphs are subjected to structural representation learning and screening by the sequential embedding mechanism, frequent subgraphs are obtained. A set of frequent subgraphs is dynamically generated. A set of higher-order hyperedges is constructed based on the set of frequent subgraphs, and a hypergraph structure is constructed based on the set of higher-order hyperedges. S2. A hypergraph convolutional network is used to perform high-order structural semantic feature aggregation and representation learning on the node features in the hypergraph structure, and to generate enhanced node representations.

[0031] The above technical solution enables the construction of a high-order structure-aware hypergraph convolutional network based on frequent subgraph mining. This solution, based on original graph data, automatically extracts high-frequency, highly representative subgraph patterns (i.e., frequent subgraphs) as high-order structural units through a data-driven approach, effectively overcoming the limitations of traditional methods that rely on manual prior knowledge. By introducing a sequential embedding mechanism to semantically rank and filter candidate subgraphs, the optimal set of subgraphs is dynamically selected as the frequent subgraph set, improving the rationality, representativeness, and task relevance of high-order structure selection. Finally, the selected frequent subgraphs are used to construct multi-type hyperedges, jointly modeling multivariate high-order relationships between nodes, significantly enhancing node representation capabilities, and thus improving the model's expressiveness, interpretability, and robustness. This invention provides a new, automated, and data-adaptive path for high-order structure modeling in the field of graph neural networks, with broad application prospects and industrialization value.

[0032] The overall scheme of this invention can be divided into two main stages: the hypergraph construction stage and the hypergraph convolution stage.

[0033] 1. Hypergraph Construction Stage: Based on the original graph structure, a frequent subgraph mining method is used to automatically identify structurally stable, semantically consistent, and representative high-order subgraph patterns from the original graph data. By introducing a sequential embedding mechanism, structural representation learning and filtering are performed on candidate subgraphs to dynamically generate a high-quality set of frequent subgraphs. This set of high-order hyperedges is then used to construct a hypergraph structure, which is finally built based on the set of high-order hyperedges.

[0034] 2. Hypergraph Convolution Stage: Based on the hypergraph structure constructed above, a hypergraph convolutional network is used to aggregate the node features of the hypergraph structure at multiple levels and high-order structural features to generate enhanced node representations for downstream classification or clustering tasks.

[0035] The hypergraph construction stage mainly involves constructing a set of high-order semantic hyperedges through frequent subgraph mining, and then constructing a hypergraph structure based on the set of high-order semantic hyperedges. The hypergraph convolution stage mainly involves performing node representation learning based on the constructed hypergraph structure. The model takes the original graph G as input and outputs the final representation of each node, thereby serving the downstream classification task.

[0036] The overall structure and construction method of the high-order structure-aware hypergraph convolutional network framework diagram in this embodiment are as follows: Figure 2 As shown.

[0037] The hypergraph construction stage and the hypergraph convolution stage will be explained in more detail below.

[0038] I. Hypergraph Construction Phase As a preferred technical solution, the step of generating candidate subgraphs based on the original graph structure through frequent subgraph mining includes: Based on the original graph structure, the neighborhood of a node is used as the basic unit. The neighborhood subgraphs within the local k-order neighborhood of a node are enumerated, and the candidate subgraph set is determined by statistically analyzing the frequency of occurrence of the neighborhood subgraphs in the whole graph.

[0039] As a preferred technical solution, the introduction of a sequential embedding mechanism to perform structural representation learning and filtering on the candidate subgraphs to obtain frequent subgraphs, and dynamically generating a set of frequent subgraphs, includes: Build node representations for each candidate subgraph and input them into a graph embedding model to learn their structural representations; The partial order distance function based on sequential embedding is used to determine the inclusion relationship between different subgraphs and whether different subgraphs constitute equivalence classes. Based on the judgment results, positive and negative sample pairs are constructed, and the graph embedding model is trained to make the embedding distance of similar subgraphs closer and the embedding distance of dissimilar subgraphs farther. Based on the embedding distance minimization criterion, the set of frequent subgraphs with the smallest error is dynamically selected from the candidate subgraph set.

[0040] This embodiment introduces a sequential embedding mechanism and graph structure similarity constraints to semantically rank and filter candidate subgraphs, dynamically selecting the optimal set of subgraphs to improve the rationality, representativeness, and task relevance of high-order structure selection.

[0041] As a preferred technical solution, the expression for the partial order distance function based on sequential embedding is:

[0042] in, The structure embedding function of the subgraph is represented. Characteristic subgraph The partial order relation between them; if Then its embedding should satisfy .

[0043] This embodiment can effectively determine the inclusion relationship between different subgraphs and whether different subgraphs constitute an equivalence class by using the partial order distance function described above.

[0044] As a preferred technical solution, the training graph embedding model adopts a discriminative ranking loss function for positive and negative sample pairs, the expression of which is:

[0045] in, Represents a pair of positive sample subgraphs that satisfy a partial order relation. For negative sample pairs that do not satisfy the partial order relation. This is the minimum interval constant between positive and negative sample pairs. Characteristic subgraph The partial order relation between them Characteristic subgraph and The partial order relationship between them.

[0046] As a preferred technical solution, the step of dynamically selecting the set of frequent subgraphs with the smallest error from the candidate subgraph set based on the embedding distance minimization criterion includes: Perform distance-compressed clustering on all embedded candidate subgraphs, and select the cluster centers representing the most embedded subgraphs based on the ordinal embedding distance. Each subgraph is a set of frequent subgraphs.

[0047] As a preferred technical solution, the step of dynamically selecting the set of frequent subgraphs with the smallest error from the candidate subgraph set based on the embedding distance minimization criterion further includes: An iterative greedy sampling strategy is adopted to gradually select the subgraph with the smallest error from the candidate subgraph set and add it to the frequent subgraph set.

[0048] In this embodiment, both distance-compressed clustering and iterative greedy sampling strategies are used to update and refine the set of frequent subgraphs. Distance-compressed clustering, through dimensionality reduction or metric learning, compresses high-dimensional, sparse data distance relationships into a low-dimensional, compact representation, significantly reducing subsequent computational complexity. Building on this, iterative greedy sampling does not involve a one-time random selection, but rather "greedily" selects the sample that best represents the data distribution or is most beneficial to model improvement in each round, dynamically adjusting the sampling strategy through multiple iterations. This combination avoids the expensive overhead of global computation while intelligently approximating the global optimum, achieving a high level of balance between computational efficiency and model accuracy.

[0049] As a preferred technical solution, the construction of a higher-order hyperedge set based on the frequent subgraph set includes: Based on the set of frequent subgraphs, for each selected frequent subgraph pattern, all matching instances of the pattern are enumerated in the original graph, and the set of nodes contained in each instance is used as a higher-order hyperedge to construct a set of higher-order hyperedges.

[0050] To make the construction method of the high-order structure-aware hypergraph convolutional network of the present invention clearer, the construction method of the high-order structure-aware hypergraph convolutional network will be further explained below in conjunction with the above-mentioned preferred technical solutions.

[0051] The hypergraph construction stage in this embodiment mainly includes three key steps: frequent subgraph mining, sequential embedding filtering, and high-order hyperedge generation, which will be described below.

[0052] 1. Frequent Subgraph Mining. Let the original simple undirected graph (i.e., the original graph) be G = (V, E), where V is the set of nodes and E is the set of edges. In this embodiment of the invention, the neighborhood of a node is used as the basic unit. Neighborhood subgraphs within the local k-order neighborhood of a node are enumerated, and the initial set of candidate subgraphs is determined by statistically analyzing the frequency of occurrence of these neighborhood subgraphs in the entire graph.

[0053] Its main algorithm flow is as follows: 1) Using each node as the center, extract its local neighborhood graph and generate a candidate set of neighborhood subgraphs; 2) Set the subgraph size k and the maximum number of candidates N, and record the frequency distribution of each type of subgraph pattern (i.e., neighborhood subgraph) in the whole graph; 3) Based on the frequency of subgraph patterns (i.e. neighborhood subgraphs) in the whole graph, select the N most frequent subgraph patterns (i.e. neighborhood subgraphs) as the initial candidate subgraph set (i.e. the initial frequent subgraph set).

[0054] 2. Sequential Embedding Filtering. To further improve the semantic consistency and structural rationality of subgraph filtering, this embodiment of the invention introduces a sequential embedding mechanism to perform structural representation and filtering of candidate subgraphs, the process of which is as follows: Figure 3 As shown. 1) Construct an ordered embedding vector for each subgraph to capture information such as the inclusion relationship and hierarchical structure of the subgraph; 2) Define a partial order distance function to measure whether subgraph A is a subpattern of subgraph B and to determine whether different subgraphs constitute an equivalence class; 3) Construct positive and negative sample pairs to train the ordered embedding model (i.e., the graph embedding model) so that the embedding distance of subgraphs of the same type is closer and the embedding distance of subgraphs of different types is farther; 4) Based on the embedding distance minimization criterion, dynamically select the frequent subgraph sample set (i.e., the frequent subgraph set) with the smallest error from the candidate subgraph set.

[0055] 3. Generation of higher-order hyperedges. For each selected frequent subgraph pattern, enumerate all matching instances of the pattern in the original graph, and use the set of nodes contained in each instance as a higher-order hyperedge to construct a set of higher-order hyperedges, ultimately forming a higher-order hypergraph structure.

[0056] To make the above hypergraph construction process clearer, the following provides a more detailed implementation process.

[0057] To fully extract the potential higher-order semantic information from the original graph structure, the method in this embodiment of the invention constructs higher-order hyperedge relationships with stronger structural interpretability based on frequent subgraph mining. The construction process mainly includes frequent subgraph mining and hyperedge generation. The frequent subgraph mining process identifies high-frequency subgraph structures in the graph and filters and embeds candidate subgraphs; hyperedge generation generates hyperedges based on these sampled structures, constructing higher-order connectivity relationships.

[0058] Let the structure of the map to be excavated (i.e., the original map structure) be... The goal of frequent subgraph identification is to identify frequently occurring subgraphs. - The node subgraph structure is recorded, along with its structural distribution and frequency of occurrence, providing a candidate set of patterns for subsequent construction of hyperedge relationships. That is, given a subgraph size... and the upper limit of the expected number of results Under the conditions, from Identify the most frequently occurring indivual Node subgraph patterns constitute the candidate set Each mode It is a picture A subgraph of .

[0059] To this end, this invention employs a structure sampling and frequency estimation mechanism based on sequential embedding distance. It constructs candidate subgraphs by combining point neighborhood structures, and estimates the frequency of each type of subgraph structure in the original graph by combining structural similarity measures in the graph embedding space (i.e. graph structure similarity constraints), and finally forms a representative subgraph pattern set (i.e. frequent subgraph set).

[0060] Specifically, firstly, the original image Decomposed into local neighborhood regions centered on each node. Using this region as the seed graph, a layer-by-layer node expansion operation is performed to generate a set of candidate subgraphs. For each subgraph... Construct its node representation and input it into the graph embedding model to learn its structural representation. .

[0061] Subsequently, based on the distance constraints between structural embeddings, it is determined whether different subgraphs constitute equivalence classes, i.e., whether they can be considered instances of the same frequent pattern. To this end, this embodiment introduces a partial order distance calculation function based on sequence embeddings:

[0062] in, The structure embedding function of the subgraph is represented. Characteristic subgraph and The partial order relation between them, if Then its embedding should satisfy Based on this, this embodiment further defines a discriminative ranking loss function for positive and negative sample pairs, optimizes the sequential embedding structure between subgraphs, and the expression for the discriminative ranking loss function for positive and negative sample pairs is as follows:

[0063] in, Represents a pair of positive sample subgraphs that satisfy a partial order relation. For negative sample pairs that do not satisfy the partial order relation. This represents the minimum margin constant between positive and negative sample pairs. After optimizing the graph embedding model using this loss function, a set of order-aware subgraph embedding representations can be obtained.

[0064] Based on this, in order to identify the most frequently occurring elements in the original image For the subgraph pattern, distance-compressed clustering needs to be performed on all embedded candidate subgraphs, and the cluster centers should be selected based on the ordinal embedding distance. Subgraphs as a set of frequent patterns Specifically, for each seed node Enumerate all sizes in its neighborhood graph Given a set of subgraph instances, compute the average ordinal distance between the embedding representation of each subgraph in the set and existing patterns:

[0065] in, Subgraph , It is a structure embedding vector (i.e., a structure embedding function representing a subgraph). The operation represents taking the non-negative part of each element to characterize the partial order constraint. The embedding process of sequential embedding in this embodiment is as follows: Figure 3 As shown.

[0066] Furthermore, this embodiment also employs an iterative greedy sampling strategy, progressively selecting the subgraph with the smallest error from the candidate set and adding it to the dataset. Let the first... In this iteration, the current set of candidate subgraphs is Select the subgraph with the smallest error from the current pattern set. :

[0067] This subgraph is the [number]th [subgraph]. Selected frequent subgraph samples. Iterate in this way until a complete set of samples is constructed. This continues until all subgraph patterns are identified. The final set of frequent subgraph patterns is denoted as:

[0068] That is, choose the one with the smallest error. Each subgraph serves as the final frequent subgraph pattern, used for subsequent hyperedge construction.

[0069] After sampling frequent subgraphs, they need to be mapped to higher-order structural units to construct the target hypergraph structure. Specifically, for each selected frequent subgraph pattern... In the original image Iterate through all the structure instances that iterates over, and construct the corresponding hyperedges. Let... Representing frequent subgraph patterns In the figure The set of matching instances in the graph, where each subgraph instance... All can be represented as a subset of nodes. Then we will each A subset of nodes contained within the set is treated as a higher-order connection unit and mapped to a hyperedge. The final set of hyperedges can be defined as follows:

[0070] in, This represents a set of higher-order hyperedges constructed by frequent subgraphs, where each hyperedge connects all nodes in an identified structural instance in the original graph.

[0071] Furthermore, based on the hyperedge set constructed above, the final hypergraph structure can be obtained:

[0072] in, Preserve the original set of nodes Furthermore, a set of higher-order hyperedges is constructed using structural instances from the frequent subgraph mining results. This allows for the explicit modeling of the relationships between diverse and complex structural patterns within the diagram.

[0073] II. Hypergraph Convolution Stage Based on the constructed high-order hypergraph structure, this embodiment of the invention uses a hypergraph convolutional neural network to learn node features.

[0074] As a preferred technical solution, the step of using a hypergraph convolutional network to perform high-order structural semantic feature aggregation and representation learning on the node features in the hypergraph structure, and generating enhanced node representations, includes the following steps: Step 1: Initialize the node features and weights in the hypergraph structure; This step is used to initialize features and weights.

[0075] Step 2: Perform feature aggregation from nodes to hyperedges and feature propagation from hyperedges back to nodes; This step involves feature propagation and convolution. The convolution process in this embodiment consists of two stages: node-to-hyperedge feature aggregation, and hyperedge-to-node feature propagation.

[0076] Step 3: Stack the above convolutional layers multiple times, with the output of each layer serving as the input of the next layer, to finally generate an enhanced node representation.

[0077] This step involves multi-layer stacked convolutions and feature output. In this embodiment, the above convolutional layers are stacked in multiple layers, with the output of each layer serving as the input to the next layer; the final node representation (i.e., the generated augmented node representation) serves as the input to the classification or clustering task.

[0078] To make the hypergraph convolution process clearer, the following provides a more detailed implementation process.

[0079] To effectively utilize the high-order structural information constructed based on subgraph mining, this embodiment introduces a hypergraph neural network on top of hypergraph construction to achieve feature aggregation and representation learning of high-order structural semantics. Unlike traditional spatial hypergraph models constructed based on the adjacency structure of the original graph, the high-order connectivity relationships proposed in this embodiment are generated driven by subgraph semantics, explicitly encoding the distribution features of local structural patterns. Based on this structural modeling approach, the model achieves feature enhancement of the full graph's structural semantics during subsequent propagation.

[0080] Specifically, let the hypergraph structure be... , record the The node features in the layer are represented as The hyperedge feature is represented as ,in Representing feature dimension, and Let represent the node set and hyperedge set of the hypergraph, respectively. In the spatial hypergraph convolution process, feature propagation includes two stages: 1) feature aggregation from nodes to hyperedges; 2) feature propagation back from hyperedges to nodes. This process can be formally defined as:

[0081]

[0082] in, Indicates the superedge The set of adjacent nodes, Represents a node The set of connected hyperedges, and These are the aggregation functions from node to hyperedge and from hyperedge back to node, respectively. To enhance the expressive power of the model, the above process can be further specified as a weighted summation:

[0083]

[0084] in, Let the degree of the superedge be . Let the degree of the node be . For nodes on the hyperedge The weights in the equation. and These are trainable linear transformation matrices; and For bias terms, This represents a non-linear activation function.

[0085] To further simplify model representation and computation, this method uses matrix form to uniformly represent the aforementioned two-stage propagation process in engineering implementation. Let the incidence matrix of the hypergraph be denoted as... The node degree matrix is The hypermarginality matrix is The convolution process can then be written as:

[0086]

[0087] in, The hyperedge weight matrix is ​​initialized as an identity matrix and can also be used as a trainable parameter to further optimize the edge weight representation. Ultimately, node features are aggregated layer by layer with their corresponding higher-order structural neighborhood features through multi-layer stacked hypergraph convolution, achieving structure-enhanced representation learning.

[0088] This mechanism can fully utilize the local high-order semantic structure revealed by frequent subgraphs, provide nodes with richer contextual information, and improve the model's structural sensitivity and semantic expressive ability.

[0089] In summary, this invention provides a method for constructing a high-order structure-aware hypergraph convolutional network based on frequent subgraph mining and its application scheme. This technical solution aims to address the problems of existing graph neural networks, such as reliance on manual priors, lack of adaptive capabilities, and difficulty in comprehensively capturing complex group interaction patterns during high-order structure modeling. Through data-driven frequent subgraph mining and sequential embedding filtering mechanisms, combined with an efficient hypergraph convolutional structure, high-order pattern modeling and node representation enhancement of complex graph data are achieved.

[0090] This invention integrates frequent subgraph mining, sequential embedding filtering, higher-order hyperedge construction, and hypergraph convolutional learning to form a complete closed-loop higher-order structure perception scheme. This scheme boasts significant advantages such as high automation, strong generalization ability, and superior node representation. It achieves efficient and high-quality higher-order graph learning without requiring manual prior conditions, effectively improving the performance of tasks such as node classification, and thus significantly outperforming existing technologies.

[0091] The specific application areas of the high-order structure-aware hypergraph convolutional network construction method of this invention include, but are not limited to, the field of civil aviation information technology.

[0092] like Figure 4 As shown, an embodiment of the present invention also provides a device for constructing a high-order structure-aware hypergraph convolutional network, the device comprising: The hypergraph construction module is used for the original graph structure. It generates candidate subgraphs through frequent subgraph mining, introduces a sequential embedding mechanism to learn the structural representation of the candidate subgraphs and filter them to obtain frequent subgraphs, dynamically generates a set of frequent subgraphs, constructs a set of higher-order hyperedges based on the set of frequent subgraphs, and constructs a hypergraph structure based on the set of higher-order hyperedges. The hypergraph convolution module is used to perform high-order structural semantic feature aggregation and representation learning on the node features in the hypergraph structure using a hypergraph convolutional network, and to generate enhanced node representations.

[0093] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0094] To further demonstrate the significant advantages and beneficial effects of the high-order structure-aware hypergraph convolutional network construction method of the present invention, experimental comparisons are provided below.

[0095] In specific experiments, the effectiveness of the technical solution of this invention was verified on classic citation network datasets (Cora, Citeseer, PubMed). As shown in Table 1, compared with traditional GCN, GAT and typical hypergraph neural networks (such as HyperGCN, HGNN+), the technical solution of this invention achieves better results in node classification accuracy.

[0096] For example, the accuracy reached 82.11% on the Cora dataset, an improvement of approximately 2.3% over HyperConv, and similar accuracy improvements were achieved on Citeseer and PubMed. These experimental results demonstrate that the technique of this invention can significantly improve the performance of graph neural networks in tasks such as node classification, possessing high practical value and promising prospects for widespread application.

[0097]

[0098] Table 1. Accuracy of node classification tasks Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing one or more programs, which, when executed, can realize the method for constructing a high-order structure-aware hypergraph convolutional network as described above.

[0099] Based on the same inventive concept, such as Figure 5 As shown, the present invention also provides an electronic device, including a processor, a communication interface, a computer-readable storage medium as described above, and a communication bus; wherein the processor, the communication interface, and the computer-readable storage medium communicate with each other through the communication bus; the processor is used to execute a program stored in the computer-readable storage medium.

[0100] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0101] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, the functional modules in the various embodiments of this invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0102] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0103] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0104] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0105] The parts not mentioned in the above embodiments are the same as or can be implemented using existing technologies, and will not be further described here.

[0106] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a high-order structure-aware hypergraph convolutional network, characterized in that, The method comprises: Based on the original graph structure, candidate subgraphs are generated by frequent subgraph mining, a sequential embedding mechanism is introduced to learn and screen the structure representation of the candidate subgraphs, and frequent subgraphs are obtained, a frequent subgraph set is dynamically generated, a high-order hyperedge set is constructed based on the frequent subgraph set, and a hypergraph structure is constructed based on the high-order hyperedge set; The node features in the hypergraph structure are aggregated and represented by a hypergraph convolution network based on high-order structure semantics, and enhanced node representations are generated.

2. The method of claim 1, wherein the method comprises: Based on the original graph structure, candidate subgraphs are generated by frequent subgraph mining, a sequential embedding mechanism is introduced to learn and screen the structure representation of the candidate subgraphs, and frequent subgraphs are obtained, a frequent subgraph set is dynamically generated, a high-order hyperedge set is constructed based on the frequent subgraph set, and a hypergraph structure is constructed based on the high-order hyperedge set; 3. The method of claim 1, wherein the method comprises: For each candidate subgraph, a node representation is constructed and input into a graph embedding model to learn its structure representation; A partial order distance function based on sequential embedding is used to determine the inclusion relationship between different subgraphs and whether different subgraphs form equivalence classes; Based on the judgment result, positive and negative sample pairs are constructed to train the graph embedding model, so that the embedding distance of the same class of subgraphs is closer and the embedding distance of different classes of subgraphs is farther; Based on the embedding distance minimization criterion, the frequent subgraph set with the minimum error is dynamically screened from the candidate subgraph set.

4. The method of claim 3, wherein the partial order distance function based on sequential embedding is expressed as:

5. The method of claim 4, wherein the training of the graph embedding model uses a discriminative ranking loss function of positive and negative sample pairs, and the function is expressed as:

6. The method of claim 3, wherein the frequent subgraph set with the minimum error is dynamically screened from the candidate subgraph set based on the embedding distance minimization criterion, and the method comprises:

7. The method of claim 6, wherein the frequent subgraph set with the minimum error is dynamically screened from the candidate subgraph set based on the embedding distance minimization criterion, and the method further comprises: wherein, denotes a structural embedding function of a subgraph, characterizes a subgraph between the partial order relation. An iterative greedy sampling strategy is used to gradually select the subgraph with the minimum error from the candidate subgraph set and add it to the frequent subgraph set.

8. The method of claim 1, wherein the high-order hyperedge set is constructed based on the frequent subgraph set, and the method comprises: wherein, denote positive sample subgraph pairs satisfying a partial order relation, denote negative sample pairs not satisfying the partial order relation, is a minimum separation constant between positive and negative sample pairs, denote a partial order relation between subgraphs denote a partial order relation between subgraphs denote a partial order relation between subgraphs denote a partial order relation between subgraphs denote a partial order relation between subgraphs ​ ​ The distance compressed clustering is performed on all embedded candidate subgraphs, and a subgraph represented by a cluster center is selected as a frequent subgraph set in the order of embedded distance as a criterion. a subgraph represented by a cluster center is selected as a frequent subgraph set in the order of embedded distance as a criterion. ​ ​ ​ ​ ​ Based on the frequent subgraph set, for each filtered frequent subgraph pattern, enumerate all matching instances of the pattern in the original graph, and take the node set contained in each instance as a high-order hyperedge to construct a high-order hyperedge set. 9.The method of claim 1-8, wherein, the method further comprises: initializing node features and weights in the hypergraph structure; performing feature aggregation from nodes to hyperedges and feature propagation from hyperedges back to nodes; stacking the convolution layers, and taking the output of each layer as the input of the next layer to finally generate enhanced node representations.

10. A high-order structure-aware hypergraph convolution network construction apparatus, characterized in that, The apparatus comprises: a hypergraph construction module configured to generate candidate subgraphs from a frequent subgraph mining on an original graph structure, introduce a sequential embedding mechanism to perform structure representation learning and filtering on the candidate subgraphs to obtain frequent subgraphs, dynamically generate a frequent subgraph set, construct a high-order hyperedge set based on the frequent subgraph set, and construct a hypergraph structure based on the high-order hyperedge set; a hypergraph convolution module configured to perform high-order structure semantic feature aggregation and representation learning on node features in the hypergraph structure by using a hypergraph convolution network, and generate enhanced node representations. 11.A computer readable storage medium storing one or more programs, wherein the one or more programs, when executed, implement the method of any one of claims 1-9. The processor, the communication interface, and the computer readable storage medium communicate with each other through a communication bus. 12.An electronic device comprising a processor, a communication interface, the computer readable storage medium of claim 11, and a communication bus; wherein, The processor is configured to execute the programs stored in the computer readable storage medium. ​ ​

Images