Article classification method and system based on improved graph structure learning
By adopting multiple sparsification patterns and adaptive graph structure construction methods in graph structure learning, the graph structure is optimized, the high computational complexity problem of large-scale graph data is solved, the graph modeling capability and classification accuracy are improved, and efficient article classification is achieved.
Patent Information
- Application Number
- CN202510895224.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2025-10-17
AI Technical Summary
Existing graph neural network methods have high computational complexity when processing large-scale graph-structured data, and sparse graph learning methods ignore graph modeling capabilities, resulting in the accidental deletion of important intra-cluster edges or the retention of noise edges, affecting classification accuracy.
By constructing a citation network training set, we use multiple sparsification modes (clustering, extended graph, original and re-weighted modes) to perform graph structure sparsification, combine adaptive graph structure construction and fusion, optimize the graph structure, use the extended graph mode to enhance the information flow between nodes, combine linear transformation and graph neural network encoding, dynamically learn edge weights, and generate an optimized graph structure.
It effectively reduces computational complexity, maintains sparsity while improving graph modeling capabilities, achieves more accurate article classification, avoids accidental deletion of important edges and retention of noise edges, and improves classification accuracy.
Smart Images
Figure CN120804337A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the technical field of graph structure processing, and in particular to an article classification method and system based on improved graph structure learning. BACKGROUND
[0002] In the era of information explosion, article data presents the characteristics of massiveness, heterogeneity and dynamics, and the traditional article classification methods based on tree structure or vector space model gradually reveal limitations. Although the tree structure is clear in hierarchy, it is difficult to handle complex cross relationships; the vector space model has deficiencies in semantic understanding and relationship expression. Under this background, the graph structure has become a research hotspot in the field of article classification due to its powerful relationship modeling ability and adaptability to complex data. The graph neural network has become a core tool in deep learning of graph structure data, and its main advantage lies in that it can combine node features and graph structure information and effectively learn the representation of nodes through an information propagation mechanism. However, many existing graph neural network methods are still based on an implicit assumption that the observed graph structure can accurately and optimally reflect the relationship between nodes. However, this assumption often does not hold in practical applications, and the reasons are mainly as follows: i) due to the inevitable errors or data noise in the data collection process, there may be false edges or missing edges in the observed graph structure. ii) the observed graph structure often only reflects the explicit connection and ignores the potential internal correlation between nodes, and these correlation relationships are crucial to the graph neural network.
[0003] To solve the above problems, graph structure learning aims to automatically optimize and learn the optimal graph structure according to the downstream task. By calculating the feature similarity between nodes or regarding the adjacency matrix as a learnable parameter matrix, graph structure learning can flexibly model the connection relationship between any node pair, thereby searching for the optimal graph structure in a wide graph topology search space. Such methods are called dense graph structure learning. However, as the number of nodes increases, the number of edges that may exist in the graph grows quadratically, so directly modeling the relationship between all node pairs requires extremely high quadratic computational complexity. This computational cost makes it difficult to generalize graph structure learning to large-scale graphs, limiting its feasibility in practical applications. To reduce the computational complexity of graph structure learning, existing research has proposed various technical approaches: 1) using low-rank approximation to avoid explicit calculation and storage of dense adjacency matrices, however, the low-rank assumption does not always hold in real-world scenarios; 2) modeling based on predefined sparse patterns, i.e., only retaining a small number of edges in the graph to reduce computational complexity. Taking the second method as an example, Chinese patent application CN113918722 provides a graph convolution clustering method based on sparse graph learning, which embeds a sparse graph structure learning layer before each graph convolution layer to construct a more optimal graph structure and improve the precision of processing the graph structure, but when performing sparse graph-based graph structure learning, the sparsification relies on data filtering, ignoring the consideration of graph modeling capability, which may lead to the deletion of important intra-cluster edges or the retention of noise edges, failing to utilize the cluster structure prior of the graph, thereby affecting the downstream task.
[0004] Therefore, it is a technical problem to provide a sparse graph-based graph structure learning method with higher accuracy. SUMMARY
[0005] The purpose of the present application is to overcome the defects of the prior art and provide an article classification method and system based on improved graph structure learning. When performing graph structure sparsification, an extended graph is combined to accelerate the flow of information between nodes, thereby effectively approximating the graph modeling capability of dense graph structure learning while maintaining sparsity, balancing the efficiency of the algorithm and the graph modeling capability.
[0006] The purpose of the present application can be achieved by the following technical solutions:
[0007] According to a first aspect of the present application, an article classification method based on improved graph structure learning is provided, the method comprising:
[0008] Constructing a citation network training set, wherein the nodes in each graph structure in the training set represent articles, and the edges represent the citation relationship between two articles;
[0009] Training an article classification model using the training set, and repeating the following steps until the performance of the article classification model is optimal:
[0010] one graph structure in the training set as an input graph, performing sparse processing on the input graph in multiple modes to obtain multiple sparse mode graphs; the multiple modes include clustering mode, extended graph mode, and original and re-weighting mode; the edges in each sparse mode graph include candidate edges;
[0011] performing adaptive graph structure construction on the graph corresponding to each sparse mode in the multiple sparse mode graphs based on the input graph;
[0012] fuse the adaptive graph structure to obtain an optimized graph structure;
[0013] input the optimized graph structure into a downstream neural network in an article classification model to generate a classification prediction, calculate a loss function based on the classification prediction, and train the article classification model based on the loss function;
[0014] use the trained article classification model to classify articles in a citation network.
[0015] As a preferred technical solution, the method of adaptive graph structure construction comprises:
[0016] perform feature encoding on the input graph to obtain node embedding, the feature encoding method comprises linear transformation encoding and graph neural network encoding, and the feature encoding method is selected depending on a gating mechanism based on homogeneity index;
[0017] based on the node embedding, calculate the edge weight of the node pair connected by the candidate edge in each of the sparse mode graphs;
[0018] optimize each of the sparse mode graphs based on the edge weight to obtain a corresponding optimized sparse mode graph;
[0019] calculate the optimized adjacency matrix of each optimized sparse mode graph, and obtain an adaptive graph structure based on the optimized adjacency matrix.
[0020] As a preferred technical solution, the method of linear transformation encoding is:
[0021]
[0022] wherein, Z L represents the node embedding of linear transformation encoding; h ω (·) represents a linear transformation encoder; X represents the input graph; M represents the number of heads; W i represents the trainable transformation matrix of the i-th head.
[0023] As a preferred technical solution, the method of graph neural network encoding uses a graph neural network comprising multiple layer structures.
[0024] computing the adjacency matrix and the degree matrix of the input features of the current layer structure in the graph neural network; in the first layer structure, the input features are the input graph, and the input features of the lth layer structure are the output features of the (l-1)th layer;
[0025] computing the output features of the current layer structure based on the adjacency matrix and the degree matrix, and the expression is: wherein, denotes the degree matrix; denotes the adjacency matrix, and A∈[0,1] N×N denotes the weighted adjacency matrix with the weight in the range of [0,1], I N denotes the N-order unit matrix; H (l) denotes the input features of the lth layer structure; W (l) denotes the layer-by-layer trainable weight matrix; σ(·) denotes a nonlinear activation function; and the output of the last layer structure is the node embedding coded by the graph neural network.
[0026] As a preferred technical solution, the gating mechanism based on the homogeneity index is:
[0027] obtaining the number of edges connecting nodes of the same type in the input graph, and computing the homogeneity based on the number;
[0028] if the homogeneity is greater than a gating threshold, selecting the graph neural network coding, otherwise selecting the linear transformation coding.
[0029] As a preferred technical solution, the method for obtaining the adaptive graph structure is:
[0030]
[0031] wherein, D (s) denotes the degree matrix of the optimized adjacency matrix S (s) corresponding to the sparsification mode s; S (s) denotes the optimized adjacency matrix corresponding to the sparsification mode s.
[0032] As a preferred technical solution, the fusion method is:
[0033]
[0034] wherein, denotes a set of sparsification modes; denotes a learnable scaling factor of the sparsification mode s normalized by a softmax function; denotes the adaptive graph structure of the sparsification mode s.
[0035] As a preferred technical solution, the normalization method is:
[0036]
[0037] Wherein, gamma (s) Indicates the learnable scaling factor of the sparsification mode s; Indicates the set of sparsification modes; gamma (i) Indicates the learnable scaling factor of the sparsification mode i.
[0038] As a preferred technical solution, the training method is: based on the loss function, the article classification model is trained end to end.
[0039] According to the second aspect of the present application, an article classification system based on improved graph structure learning is provided for implementing the above method.
[0040] Compared with the prior art, the present application sparsifies the graph structure of the citation network in multiple modes based on the extended graph, only retains the intra-cluster edges of the citation network graph structure through the clustering mode, removes the inter-cluster edges, realizes structure simplification, avoids absolute dependence on numerical values, introduces extended graph mode sparsification on this basis to solve the cross-cluster association fragmentation caused by the clustering mode, and enhances the connection strength between nodes. Finally, the original and re-weighted modes are used to eliminate bias. Through the above technical means, the important edge deletion error and noise edge retention problem of the graph structure can be effectively avoided, the sparsity of the graph structure is maintained, and the information flow of the nodes in the graph structure is accelerated, so that more accurate and efficient article classification is realized. BRIEF DESCRIPTION OF DRAWINGS
[0041] Figure 1 The method flowchart of the present application is shown. DETAILED DESCRIPTION
[0042] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.
[0043] Unless otherwise defined, technical terms and scientific terms used in the present application shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terms "a", "an", "one", "this", and similar referents in the context of describing the application are to be construed to be inclusive, not exclusive. The terms "comprise", "comprising", "comprises", "include", "including", "includes", "contain", "containing", "contains", and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, product, or apparatus that comprises a list of steps or units does not necessarily comprise only those steps or units but can include additional steps or units not expressly listed or inherent to such process, method, product, or apparatus. The terms "connected", "coupled", "linking", and similar referents in the context of this application are to be construed as not necessarily referring to a direct or indirect physical or mechanical connection, but can include an electrical connection, whether direct or indirect. The term "multiple" refers to two or more. The term "and / or" describes associated objects in association with the associated objects, which means that there are three relationships, for example, "A and / or B" can mean that A exists alone, A and B exist together, and B exists alone. The character " / " generally represents an "or" relationship between the associated objects. The terms "first", "second", "third", and the like only distinguish similar objects, and do not represent a specific order of the objects.
[0044] To solve the problems existing in the prior art, the application provides an article classification method based on improved graph structure learning, and the flowchart is as shown in Figure 1
[0045] In detail, it includes:
[0046] S1, constructing a citation network training set, wherein the nodes in each graph structure in the training set represent articles, and the edges represent the citation relationship between two articles.
[0047] Training the article classification model using the training set, and repeating the following steps until the performance of the article classification model is optimal:
[0048] S2, taking the graph structure in the training set as an input graph, and performing sparse processing on the input graph in multiple modes to obtain multiple sparse mode graphs.
[0049] Specifically, the multiple modes include clustering mode, extended graph mode, and original and reweighting mode, and the details of the above three modes are as follows:
[0050] 1) Clustering mode:
[0051] Due to the natural tendency of more tightly connected within clusters, real-world graphs are usually composed of multiple tightly connected clusters, while the connections between clusters are much sparser and less likely to establish new connections. Based on this, the present application is based on the attribute of centralized edge distribution, and a clustering mode is designed, which retains the basic connection within the cluster and removes the inter-cluster edge to achieve sparsity. Specifically, in order to identify potential clusters, an efficient graph partitioning algorithm such as METIS is used to partition the input graph into multiple disjoint clusters, and all possible connections within each cluster are considered as candidate edges of the clustering mode. In this way, the inherent structural attribute based on centralized edge distribution is used to ensure the retention of basic connection.
[0052] 2), extended graph mode:
[0053] The present application considers that the dense graph structure learning is the most expressive model, because any two nodes in the dense graph can be explicitly connected. In contrast, the sparse mode only contains partial connections, so an ideal sparse mode should approximate the connectivity of the dense graph structure through multi-hop propagation as an alternative. In other words, the sparse mode should allow each node to reach any other node within multi-hop propagation, so that the generated sparse graph implicitly captures all possible interactions after several rounds of message passing using downstream graph convolutional networks, as shown in the upper right corner. More importantly, the convergence from sparse to dense should be accelerated, so that the connectivity approximation can be efficiently completed with fewer layers in the downstream graph neural network. Figure 1
[0054] Although the clustering mode captures the basic connection and provides a stable foundation, due to its limited flexibility, it cannot capture fine-grained connections, so another form is needed to supplement it. In the present application, the sparse mode based on the extended graph is selected to expand the search space of the graph structure and enhance the graph modeling capability on the basis of the basic clustering mode. To achieve this goal, the present application proposes an extended graph mode based on the extended graph, which can efficiently approximate the connectivity of the dense graph structure learning. Specifically, the expanded graph has the following connection characteristics: the shortest path length between any two nodes will increase logarithmically with the increase of the number of nodes. This means that when the size of the graph expands, the growth rate of the required propagation iteration will gradually slow down, thereby accelerating the flow of information between nodes. Specifically, the present application constructs an extended graph based on the mainstream extended graph construction method, and considers all edges of the extended graph as candidate edges of the extended graph mode.
[0055] 3), original and re-weighted mode:
[0056] In order to emphasize the key connection while filtering out the noise edges in the input graph, in addition to the clustering and extended graph modes, the present application also introduces the original and re-weighted mode for sparsification.
[0057] Specifically, in this mode, all edges in the input graph are regarded as candidate edges of the sparsification mode, and the sparsification mode in which edges are not allowed to be reweighted is recorded as an original mode, in which the weights of all edges are fixed as 1; the sparsification mode in which edges are allowed to be reweighted is recorded as a reweighting mode.
[0058] S3, adaptively constructing a graph corresponding to each sparsification mode in the multi-sparsification mode graph based on the input graph.
[0059] S31, obtaining node embedding by performing feature encoding on the input graph, the feature encoding method including linear transformation encoding and graph neural network encoding, and the feature encoding method selection depends on a gating mechanism based on a homogeneity index.
[0060] When performing encoding method selection, for the graph neural network encoder, the application adopts the multi-layer graph convolutional neural network described in the problem definition part, which encodes the graph structure into node embedding, however, it assumes that there is strong homogeneity in the input graph, while the linear transformation encoder does not make this assumption. This prompts the application to adopt a gating mechanism based on a homogeneity index to automatically select a suitable encoder from the two candidate encoders.
[0061] wherein the linear transformation encoding method is:
[0062]
[0063] wherein Z L represents the node embedding of linear transformation encoding; h ω (·) represents the linear transformation encoder; X represents the input graph; M represents the number of heads; W i represents the trainable transformation matrix of the i-th head.
[0064] The method for performing graph neural network encoding by using the graph neural network comprising a plurality of layer structures is:
[0065] An adjacency matrix and a degree matrix of the input feature of the current layer structure in the graph neural network are calculated; in the first layer structure, the input feature is the input graph, the input feature of the l-th layer structure is the output feature of the (l-1)-th layer, and the output feature of the current layer structure is calculated based on the adjacency matrix and the degree matrix, and its expression is:
[0066]
[0067] wherein, represents the degree matrix; represents the adjacency matrix, and A∈[0,1] N×N represents the weighted adjacency matrix with the weight in the range of [0, 1], I N represents an N-order unit matrix; H(l) input features representing the l-th layer structure; W (l) representing a layer-wise trainable weight matrix; σ(·) represents a nonlinear activation function; and the output of the last layer structure is the node embedding encoded by the graph neural network.
[0068] Specifically, for a dataset whose homogeneity exceeds a gating threshold, the application uses a graph neural network encoder; otherwise, the application uses a linear transformation encoder, wherein the calculation method of homogeneity comprises: obtaining the number of edges connecting nodes of the same type in the input graph, and dividing the number by the total number of edges in the input graph to obtain the homogeneity.
[0069] S32, based on the node embedding, calculating the edge weight of the node pair connected by the candidate edge in each sparse pattern graph.
[0070] S33, optimizing each sparse pattern graph based on the edge weight to obtain the corresponding optimized sparse pattern graph.
[0071] S34, calculating the feature similarity matrix of each optimized sparse pattern graph using cosine similarity, setting the edges in the feature similarity matrix smaller than the preset value ∈ to 0 to obtain the optimized adjacency matrix, and obtaining the adaptive graph structure based on the optimized adjacency matrix.
[0072] In detail, the expression of obtaining the adaptive graph structure is:
[0073]
[0074] wherein, D (s) represents the degree matrix of the optimized adjacency matrix S (s) corresponding to the sparse pattern s; S (s) represents the optimized adjacency matrix corresponding to the sparse pattern s.
[0075] S4, fusing the adaptive graph structure to obtain the optimized graph structure.
[0076] S5, inputting the optimized graph structure into the downstream neural network in the article classification model to generate a classification prediction, calculating a loss function based on the classification prediction, and performing end-to-end training on the article classification model based on the loss function.
[0077] To further enhance the flexibility of the graph learner, the application designs an adaptive graph structure fusion mechanism, which dynamically learns the attention weight of each sparse pattern, and the calculation method of the finally optimized graph structure A * is as follows:
[0078]
[0079] wherein, represents a set of sparse patterns; denotes a learnable scaling factor for the sparse pattern s normalized by a softmax function; denotes an adaptive graph structure for the sparse pattern s.
[0080] wherein the normalization method is:
[0081]
[0082] wherein γ (s) denotes a learnable scaling factor for the sparse pattern s; denotes a set of sparse patterns; γ (i) denotes a learnable scaling factor for the sparse pattern i.
[0083] S6, classifying the citation network article by using the trained article classification model.
[0084] S61, inputting the citation network graph structure to be classified into the trained article classification model.
[0085] S62, sparse the citation network graph structure to be classified including clustering mode, extended graph mode and original and re-weighting mode.
[0086] S63, construct an adaptive graph structure for the sparse graph after each mode is sparse, and re-assign edge weights for the allowed edge weight in the re-weighting mode.
[0087] S64, adaptively fuse all sparse graphs to obtain an optimized graph structure as an input of a downstream graph neural network, and generate a corresponding article classification prediction, and the expression is:
[0088]
[0089] wherein A * denotes an optimized graph structure; X denotes a node feature matrix corresponding to the citation network graph structure to be classified; f θ (·) denotes a graph neural network processing.
[0090] And in the present application, the whole process of constructing the optimized graph structure and generating the prediction is differentiable.
[0091] In addition, the present application also provides an article classification system based on improved graph structure learning, which includes a central processing unit (CPU) which can perform various appropriate actions and processing according to computer program instructions stored in a read-only memory (ROM) or computer program instructions loaded into a random access memory (RAM) from a storage unit. In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM and RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.
[0092] The plurality of components in the device are connected to the I / O interface, including: an input unit, such as a keyboard, a mouse, etc.; an output unit, such as various types of displays, speakers, etc.; a storage unit, such as a magnetic disk, an optical disk, etc.; and a communication unit, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit allows the device to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunication networks.
[0093] The processing unit performs the various methods and processes described above, such as the methods S1-S6. For example, in some embodiments, the methods S1-S6 can be implemented as a computer software program, which is tangibly embodied in a machine-readable medium, such as the storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device via the ROM and / or the communication unit. When the computer program is loaded onto the RAM and executed by the CPU, one or more steps of the methods S1-S6 described above can be performed. Alternatively, in other embodiments, the CPU can be configured to perform the methods S1-S6 by any other suitable means, such as by means of firmware.
[0094] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, and without limitation, illustrative types of hardware logic components that can be used include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip systems (SOCs), complex programmable logic devices (CPLDs), etc.
[0095] Program code for carrying out methods of the present application can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the program code, when executed by the processor or controller, causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can execute entirely on a machine, partly on the machine, as a stand-alone software package, partly on the machine and partly on a remote machine or entirely on the remote machine or server.
[0096] In the context of the present application, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable storage medium can include, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage medium would include one or more lines of electrical wire, portable computer diskette, hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the foregoing.
[0097] The above description is only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. An article classification method based on improved graph structure learning, characterized in that: The method includes: Construct a citation network training set, where each node in the graph structure represents an article and an edge represents the citation relationship between two articles; The article classification model is trained using the training set, and the following steps are repeated until the article classification model has the best performance: Using a graph structure in a training set as an input graph, performing multiple pattern sparsification on the input graph to obtain multiple sparse pattern graphs; the multiple patterns include a clustering pattern, an expanded graph pattern, and an original and reweighted pattern; in each sparse pattern graph, edges include candidate edges; Based on the input graph, adaptively construct a graph structure corresponding to each sparsification pattern in the multiple sparsity pattern graph; Fusing the adaptive graph structures to obtain an optimized graph structure; Inputting the optimized graph structure into a downstream neural network in an article classification model to generate classification predictions, calculating a loss function based on the classification predictions, and training the article classification model based on the loss function; Use the trained article classification model to classify citation network articles.
2. The article classification method based on improved graph structure learning according to claim 1, characterized in that: The method for constructing an adaptive graph structure includes: Performing feature encoding on the input graph to obtain node embeddings, wherein the feature encoding method includes linear transformation encoding and graph neural network encoding, and the feature encoding method selection depends on a gating mechanism based on a homogeneity indicator; calculating edge weights for each pair of nodes connected by a candidate edge in the sparse pattern graph based on the node embeddings; Optimizing each of the sparse pattern graphs based on the edge weights to obtain a corresponding optimized sparse pattern graph; An optimized adjacency matrix of each optimized sparse pattern graph is calculated, and an adaptive graph structure is obtained based on the optimized adjacency matrix.
3. The article classification method based on improved graph structure learning according to claim 2, characterized in that: The linear transform coding method is: Among them, Z L represents the node embedding of linear transform coding; h ω (·) represents the linear transform encoder; X represents the input image; M represents the number of heads; W i represents the trainable transformation matrix of the i-th head.
4. The article classification method based on improved graph structure learning according to claim 2, characterized in that: The method for performing the graph neural network encoding using a graph neural network comprising a multi-layer structure is as follows: Compute the adjacency matrix and degree matrix of the input features of the current layer structure in the graph neural network; in the first layer structure, the input features are the input graph, and the input features of the lth layer structure are the output features of the l-1th layer; The output features of the current layer structure are calculated based on the adjacency matrix and degree matrix, and the expression is: in, represents the degree matrix; represents the adjacency matrix, and A∈[0,1] N×N represents the weighted adjacency matrix with weights in the range [0,1], I N represents the N-order unit matrix; H (l) Represents the input features of the l-th layer structure; W (l) represents a layer-by-layer trainable weight matrix; σ(·) represents a nonlinear activation function; and the output of the last layer structure is the node embedding encoded by the graph neural network.
5. The article classification method based on improved graph structure learning according to claim 2, characterized in that: The gating mechanism based on homogeneity index is: Obtaining the number of edges connecting nodes of the same type in the input graph, and calculating homogeneity based on the number; If the homogeneity is greater than the gate threshold, graph neural network encoding is selected, otherwise linear transformation encoding is selected.
6. The article classification method based on improved graph structure learning according to claim 2, characterized in that: The method for obtaining the adaptive graph structure is: Among them, D (s) Represents the optimized adjacency matrix S corresponding to the sparsification pattern s (s) degree matrix; S (s) represents the optimized adjacency matrix corresponding to the sparsification pattern s.
7. The article classification method based on improved graph structure learning according to claim 1, characterized in that: The fusion method is: in, represents a collection of sparsification patterns; represents the learnable scaling factor of the sparse pattern s normalized by the softmax function; An adaptive graph structure representing the sparsification pattern s.
8. The article classification method based on improved graph structure learning according to claim 7, characterized in that: The normalization method is: Among them, γ (s) represents the learnable scaling factor of the sparsification pattern s; represents the set of sparsification patterns; γ (i) represents the learnable scaling factor for the sparsification pattern i.
9. The article classification method based on improved graph structure learning according to claim 1, characterized in that: The training method is: performing end-to-end training on the article classification model based on the loss function.
10. An article classification system based on improved graph structure learning, characterized in that: The system is used to implement the method according to any one of claims 1 to 9.