Data classification method, device, equipment and storage medium based on linear feature enhancement

Abstract

The present invention discloses a data classification method, apparatus, device, and storage medium based on linear feature enhancement. The present invention obtains a first linear feature and a second linear feature by linearly propagating node features and corresponding neighbor node features, respectively; nonlinearly mapping the first linear feature and the second linear feature using KAN nonlinear mapping to obtain a first mapping node feature and a second mapping node feature; determining the embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature; superimposing the embeddings of each layer to obtain a target node feature representation of the first node; determining an enhanced feature of the first node based on a Sigmoid activation function, an activation function, and the target node feature representation; and converting the enhanced feature using a softmax function to obtain a classification result. The method effectively ensures the computational efficiency of graph data classification while improving the classification capability of complex graph data.

Description

Technical Field

[0001] The present invention relates to the field of data classification technology, and in particular to a data classification method, device, equipment and storage medium based on linear feature enhancement. Background Art

[0002] Existing techniques for classifying medical graph datasets typically use nonlinear models to represent complex patterns in the data. However, while nonlinear models offer significant advantages in capturing complex relationships, they also increase computational complexity and training time. Improving the model's ability to classify complex graph data while maintaining computational efficiency has become a key research challenge. Summary of the Invention

[0003] Based on this, it is necessary to address the above problems and propose a data classification method, device, equipment and storage medium based on linear feature enhancement.

[0004] A data classification method based on linear feature enhancement is used to classify graph data. The graph data contains multiple nodes, each node has a neighbor node, each node corresponds to a node feature, and each neighbor node corresponds to a neighbor node feature. The method includes:

[0005] Performing linear propagation on the node feature and the corresponding neighbor node feature to obtain a first linear feature and a second linear feature; the node corresponding to the node feature is the first node;

[0006] A first mapping node feature and a second mapping node feature are obtained by performing nonlinear mapping on the first linear feature and the second linear feature respectively through KAN nonlinear mapping;

[0007] Determining an embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through a graph convolutional network;

[0008] Superimposing the embeddings of each layer to obtain a target node feature representation of the first node;

[0009] Based on the Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node;

[0010] The enhanced features are transformed by a softmax function to obtain a classification result.

[0011] In one embodiment, linearly propagating the node feature and the corresponding neighbor node feature to obtain the first linear feature and the second linear feature is implemented by the following expression:

[0012]

[0013]

[0014] in, is the node feature, is the neighbor node feature, is the first linear feature, is the second linear feature, is the adjacency matrix, is the power of the adjacency matrix.

[0015] In one embodiment, the first mapping node feature and the second mapping node feature obtained by nonlinearly mapping the first linear feature and the second linear feature respectively through the KAN nonlinear mapping are implemented by the following expressions:

[0016]

[0017]

[0018] in, is the first mapping node feature, is the second mapping node feature, and is the trainable coefficient, is the first linear feature, is the second linear feature, is the grid number of the Fourier basis, is the frequency index.

[0019] In one embodiment, the determining of the embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through the graph convolutional network is implemented by the following expression:

[0020]

[0021] in, Represented as a node In the Layer embedding, is the first mapping node feature, is the second mapping node feature, Gather for neighbors, To transfer the node-based KAN activation function, is element-wise multiplication.

[0022] In one embodiment, superimposing the embeddings of each layer to obtain the target node feature representation of the first node is achieved by the following expression:

[0023]

[0024] in, is the target node feature representation of the first node u, 、 ,..., Represent the features of the first node u at different layers, is the total number of layers, Layer is one of the layers.

[0025] In one embodiment, the Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node through the following expression:

[0026]

[0027]

[0028] =

[0029] in, is the weighted feature, is the first mapping node feature, is the adaptive weight, is the Sigmoid activation function, is the first linear transformation, is the second linear transformation, is the activation function, is the third mapping node feature represented by the target node feature, is the target node feature representation of the first node u, and is the trainable coefficient, is the grid number of the Fourier basis, is the frequency index.

[0030] In one embodiment, the conversion of the enhanced features by the softmax function to obtain the classification result is achieved by the following expression:

[0031]

[0032] in, is the classification result, is the weight matrix, The third mapping node feature is represented by the target node feature.

[0033] A data classification device based on linear feature enhancement, the device comprising:

[0034] a propagation module, configured to perform linear propagation on the node feature and the corresponding neighbor node feature respectively to obtain a first linear feature and a second linear feature; the node corresponding to the node feature is the first node;

[0035] A mapping module, configured to perform nonlinear mapping on the first linear feature and the second linear feature respectively through KAN nonlinear mapping to obtain a first mapping node feature and a second mapping node feature;

[0036] a combining module, configured to determine an embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through a graph convolutional network;

[0037] a superposition module, configured to superimpose the embeddings of each layer to obtain a target node feature representation of the first node;

[0038] Determine the module for Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node;

[0039] An acquisition module is used to convert the enhanced features through a softmax function to obtain a classification result.

[0040] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0041] Performing linear propagation on the node feature and the corresponding neighbor node feature to obtain a first linear feature and a second linear feature; the node corresponding to the node feature is the first node;

[0042] A first mapping node feature and a second mapping node feature are obtained by performing nonlinear mapping on the first linear feature and the second linear feature respectively through KAN nonlinear mapping;

[0043] Determining an embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through a graph convolutional network;

[0044] Superimposing the embeddings of each layer to obtain a target node feature representation of the first node;

[0045] Based on the Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node;

[0046] The enhanced features are transformed by a softmax function to obtain a classification result.

[0047] A computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0048] Performing linear propagation on the node feature and the corresponding neighbor node feature to obtain a first linear feature and a second linear feature; the node corresponding to the node feature is the first node;

[0049] A first mapping node feature and a second mapping node feature are obtained by performing nonlinear mapping on the first linear feature and the second linear feature respectively through KAN nonlinear mapping;

[0050] Determining an embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through a graph convolutional network;

[0051] Superimposing the embeddings of each layer to obtain a target node feature representation of the first node;

[0052] Based on the Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node;

[0053] The enhanced features are transformed by a softmax function to obtain a classification result.

[0054] The present application obtains a first linear feature and a second linear feature by performing linear propagation on the node feature and the corresponding neighbor node feature respectively; the node corresponding to the node feature is the first node; the first linear feature and the second linear feature are nonlinearly mapped by KAN nonlinear mapping to obtain a first mapping node feature and a second mapping node feature; the embedding of each layer of the first node in the graph convolution network is determined by combining the first mapping node feature and the second mapping node feature through a graph convolution network; the embedding of each layer is superimposed to obtain the target node feature representation of the first node; based on the Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node; the enhanced features are transformed by the softmax function to obtain the classification result. This effectively ensures the computational efficiency of graph data classification while improving the classification ability of complex graph data. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0056] in:

[0057] Figure 1 This is a diagram of an application environment of a data classification method based on linear feature enhancement in one embodiment;

[0058] Figure 2 is a flow chart of a data classification method based on linear feature enhancement in one embodiment;

[0059] Figure 3 is a structural block diagram of a data classification device based on linear feature enhancement in one embodiment;

[0060] Figure 4 A comparison chart showing the impact of the positional relationship between the linear layer and the KAN layer on accuracy in one embodiment;

[0061] Figure 5 A comparison chart of the running time of KSGC and SGC in one embodiment;

[0062] Figure 6 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0064] In the prior art, when classifying graph datasets in the medical field, nonlinear models are usually used to express the complex patterns of the data. However, although nonlinear models have obvious advantages in capturing complex relationships, they also bring about the problem of increased computational complexity and correspondingly longer training time. How to improve the model's ability to classify complex graph data while ensuring computational efficiency has become an important issue in current research. In order to solve the above technical problems, the present application provides a data classification method based on linear feature enhancement.

[0065] Figure 1 FIG. 1 is an application environment diagram of a data classification method based on linear feature enhancement in one embodiment. Figure 1 , the data classification method based on linear feature enhancement is applied to a data classification system based on linear feature enhancement. The data classification system based on linear feature enhancement includes a terminal 110 and a server 120. The terminal 110 and the server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal. The mobile terminal can be at least one of a mobile phone, a tablet computer, a laptop computer, etc. The server 120 can be implemented as an independent server or a server cluster composed of multiple servers. The terminal 110 is used to perform linear propagation on the node features and the corresponding neighbor node features respectively to obtain the first linear feature and the second linear feature, and the server 120 is used to perform nonlinear mapping on the first linear feature and the second linear feature respectively through KAN nonlinear mapping to obtain the first mapping node feature and the second mapping node feature; determine the embedding of the first node in each layer of the graph convolution network by combining the first mapping node feature and the second mapping node feature through a graph convolution network; superimpose the embedding of each layer to obtain the target node feature representation of the first node; based on the Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node; the enhanced features are converted by a softmax function to obtain a classification result.

[0066] like Figure 2 As shown, in one embodiment, a data classification method based on linear feature enhancement is provided. The method can be applied to both terminals and servers. This embodiment is described by applying it to terminals. The data classification method based on linear feature enhancement is used to classify graph data in the medical field. The graph data contains multiple nodes, each node has a neighbor node, each node corresponds to a node feature, and each neighbor node corresponds to a neighbor node feature, such as Figure 2 As shown, the method includes:

[0067] S10: performing linear propagation on the node feature and the corresponding neighbor node feature to obtain a first linear feature and a second linear feature; the node corresponding to the node feature is the first node;

[0068] S20: performing nonlinear mapping on the first linear feature and the second linear feature respectively through KAN nonlinear mapping to obtain a first mapping node feature and a second mapping node feature;

[0069] S30: Determine the embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through a graph convolutional network;

[0070] S40: superimposing the embeddings of each layer to obtain a target node feature representation of the first node;

[0071] S50: Based on Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node;

[0072] S60: Convert the enhanced features through a softmax function to obtain a classification result.

[0073] The present application obtains a first linear feature and a second linear feature by performing linear propagation on the node feature and the corresponding neighbor node feature respectively; the node corresponding to the node feature is the first node; the first linear feature and the second linear feature are nonlinearly mapped by KAN nonlinear mapping to obtain a first mapping node feature and a second mapping node feature; the embedding of each layer of the first node in the graph convolution network is determined by combining the first mapping node feature and the second mapping node feature through a graph convolution network; the embedding of each layer is superimposed to obtain the target node feature representation of the first node; based on the Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node; the enhanced features are transformed by the softmax function to obtain the classification result. This effectively ensures the computational efficiency of graph data classification while improving the classification ability of complex graph data.

[0074] In one embodiment, the linear propagation of the node feature and the corresponding neighbor node feature in step S10 to obtain the first linear feature and the second linear feature is implemented by the following expression:

[0075] (1)

[0076] (2)

[0077] in, is the node feature, is the neighbor node feature, is the first linear feature, is the second linear feature, is the adjacency matrix, is the power of the adjacency matrix.

[0078] Specifically, after adding the self-loop, the spectralization process can avoid the interference of high-frequency noise on the node features and the corresponding neighbor node features. The low-pass filter is used to adjust the node features and the corresponding neighbor node features in the graph data.

[0079] In one embodiment, the first mapping node feature and the second mapping node feature obtained by nonlinearly mapping the first linear feature and the second linear feature respectively through KAN nonlinear mapping in step S20 are implemented by the following expressions:

[0080] (3)

[0081] (4)

[0082] in, is the first mapping node feature, is the second mapping node feature, and is the trainable coefficient, is the first linear feature, is the second linear feature, is the grid number of the Fourier basis, is the frequency index.

[0083] Specifically, the Kolmogorov–Arnold network (KAN) is introduced to perform nonlinear mapping on the first mapping node features and the second mapping node features. KAN guarantees the ability to approximate any continuous multivariate function. This application designs a KAN mapping layer to replace the linear transformation of the traditional neural network with a learnable unary function. To solve this problem, this paper proposes to reconstruct the feature extraction process by transferring the node basis enhancement of nonlinear mapping. The core idea of ​​the KAN design is to use cosines of different frequencies to and sine The linear combination of the basis functions can fully approximate complex nonlinear functions. For example, when When , the basis function captures the low-frequency characteristics. and The weight of each channel is used to dynamically adjust the feature expression ability of each channel.

[0084] In one embodiment, the step S30 of determining the embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through the graph convolutional network is implemented by the following expression:

[0085] (5)

[0086] in, Represented as a node In the Layer embedding, is the first mapping node feature, is the second mapping node feature, Gather for neighbors, To transfer the node-based KAN activation function, is element-wise multiplication.

[0087] Specifically, the first mapping node feature and the second mapping node feature obtained by KAN nonlinear mapping are integrated into the SGC framework (Simplified Graph Convolution Network), as shown in Formula (5). This formula enhances the expression ability of information by transferring the nonlinear transformation of the node basis: The first term represents the node The first mapping node feature and its neighbor nodes The second mapping node feature (self-product) is combined to form the basis of information propagation; the second term represents the node and neighbor nodes The interaction between them.

[0088] In one embodiment, the step S40 of superimposing the embeddings of each layer to obtain the target node feature representation of the first node is implemented by the following expression:

[0089] (6)

[0090] in, is the target node feature representation of the first node u, 、 ,..., Represent the features of the first node u at different layers, is the total number of layers, Layer is one of the layers.

[0091] Specifically, this design retains the lightweight architecture of SGC, while significantly improving the model's adaptability to graph data by introducing nonlinear capabilities through the transfer node-based KAN.

[0092] In one embodiment, for the Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node through the following expression:

[0093] (7)

[0094] (8)

[0095] = (9)

[0096] in, is the weighted feature, is the first mapping node feature, is the adaptive weight, is the Sigmoid activation function, is the first linear transformation, is the second linear transformation, is the activation function, is the third mapping node feature represented by the target node feature, is the target node feature representation of the first node u, and is the trainable coefficient, is the grid number of the Fourier basis, is the frequency index.

[0097] In one embodiment, the conversion of the enhanced features by the softmax function to obtain the classification result is achieved by the following expression:

[0098] (10)

[0099] in, is the classification result, is the weight matrix, The third mapping node feature represents the target node feature. SGC avoids the introduction of nonlinear activation functions, significantly improving computational efficiency and making it suitable for processing large-scale graph data. In this way, this application can process complex graph data, maintaining efficient computation while effectively improving the model's expressive power, overcoming the shortcomings of simplified graph convolutional networks.

[0100] Compared to traditional models that combine affine transformations with activation functions, the KAN-based network architecture combined with SGC breaks the limitations of linear weights, allowing each dimension of features to be adjusted by a specially trained nonlinear function without disrupting the original simplified linear structure of the model. This gives the model greater expressive power, enabling it to capture highly nonlinear interactions between neighboring features and its own features. In this method, the node representation after KAN mapping, denoted as Z = KAN(H), serves as the input representation for downstream tasks such as node classification. The introduction of the KAN module provides the model with a solid theoretical foundation and expressive potential, overcoming the shortcomings of the linear propagation of SGC.

[0101] In summary, this method ensures good performance on complex graph data, especially when processing large-scale data. L2 regularization effectively constrains the complexity of the model, message dropping enhances robustness, and the low-pass filter suppresses high-frequency noise, ensuring an efficient training process. This invention is verified by the following:

[0102] In the experimental validation section, we evaluated the Fourier transform-based nonlinear feature enhancement and dynamic filtering module. We utilized the commonly used graph neural network datasets: Cora, Pubmed, and Citeseer, and designed validation and comparative experiments. We used ACC and AdaBoost as the primary evaluation metrics to analyze and verify the effectiveness of this chapter's method in clustering high-dimensional data.

[0103] Several graph neural network models are compared with traditional SGC methods. First, GCN performs convolution operations on graph structures using node neighborhood information and captures high-order relationships between nodes in the graph through multi-layer graph convolution. GAT introduces a self-attention mechanism, allowing each node to assign different weights to its neighbors based on their importance when calculating their influence. FastGCN is an improved GCN method that reduces the computational complexity of graph convolution operations by approximating the graph convolution process to a sampling method, thereby improving the training efficiency of graph neural networks. GIN further improves the ability to learn graph structures by introducing a graph isomorphic network structure with strong representation capabilities. LNet is a graph neural network model that incorporates local information in the graph and improves graph representation capabilities by modeling the local structure of nodes. AdaLNet further enhances the learning of important information in the graph by incorporating an adaptive weighting mechanism, thereby improving the ability to process complex graph structures. DGI utilizes both local and global information in the graph for training, optimizing the graph representation capabilities and enhancing the graph embedding effect. Finally, SGC simplifies the convolutional computation of traditional GCNs by directly applying a linear transformation to reduce computational effort, improving efficiency while preserving the graph's structural information. The introduction of these methods provides a variety of benchmarks for model comparison and optimization. While traditional SGC sacrifices some representational power compared to complex network models like GAT and GIN, its simplified computation reduces the risk of overfitting and avoids training issues caused by high-frequency components. SGC's simplicity makes it easier to debug and deploy, making it suitable for tasks requiring high computational efficiency. Despite this limitation, SGC may not be able to compete with more advanced graph neural network methods (such as GAT and GIN) on complex tasks due to its simple structure. However, its high efficiency and low complexity make it an attractive option for certain application scenarios. The method in this chapter takes all the above-mentioned GCN, GAT, FastGCN, GIN, LNet, DGI and SGC algorithms as comparison baselines, and uses two knowledge enhancement modules, Fourier basis enhanced nonlinear mapping and channel adaptive focusing unit, to perform node classification tasks under different settings. The two evaluation indicators ACC and ADABOOST are used to evaluate the excellence of the method to prove the superiority of the method in this chapter.

[0104] The feature extraction method used is as follows:

[0105] Fourier Transform: The KSGC algorithm (present invention) uses the Fourier transform to perform nonlinear enhancement on the input node features and their corresponding neighboring node features. The Fourier transform converts the input signal from the time domain to the frequency domain, enabling the algorithm to extract more useful information from frequency features. The Fourier transform helps capture the frequency characteristics of the data and enhances the network's ability to learn local features of the data.

[0106] Adaptive Focusing: KSGC combines CAFU, a module that adaptively adjusts the focus of features through weights to further optimize the feature extraction process. The adaptive focusing module can dynamically enhance certain key features while suppressing unimportant information during the feature learning process.

[0107] Graph Convolution: During feature extraction, the KSGC algorithm further enhances feature representation with graph convolution layers. Graph convolution leverages graph structural information for feature learning, enabling the model to effectively capture the relationships and structural information between graph nodes.

[0108] Feature Enhancement: Through its feature enhancement mechanism, KSGC strengthens the expressiveness of input features and enhances the model's robustness to data by introducing more feature information. This enhancement mechanism combines the aforementioned Fourier transform with graph convolutional networks to improve the depth and precision of data representation.

[0109] Based on the experimental results in 1, we can conclude that KSGC achieves higher node classification accuracy than other models. Experimental results for each method on the Cora, PubMed, and Citeseer datasets are compared with the classification accuracy of KSGC and the baseline model on these three datasets. KSGC outperforms other models on Cora, Citeseer, and PubMed.

[0110]

[0111] The ablation results in Table 2 provide the following conclusions. Ablation experiments validate the contributions of the GraphKAN and CAFU modules. Removing GraphKAN further degrades the accuracy of PubMed, Cora, and Citeseer, as does removing CAFU. This demonstrates that GraphKAN enhances neighborhood feature interactions through Fourier kernels, while CAFU optimizes feature fusion through channel-wise attention; both are essential.

[0112]

[0113] The experimental results in Table 3 indicate that SGC + AdaBoost performs significantly worse than standalone SGC on the test set, and the training ACC is much higher than the test ACC, indicating severe overfitting. This is directly related to the conflict between SGC's linear characteristics and AdaBoost's nonlinear integration goal, further demonstrating SGC's linear model advantage. Figure 1 This article demonstrates the performance impact of the GraphKAN layer before and after the linear layer. When GraphKAN is placed before the linear layer, the ACC for PubMed, Cora, and Citeseer all improve. This is because GraphKAN's Fourier kernel transform a priori enhances the frequency domain information of the input features, providing a more discriminative feature basis for subsequent linear propagation. Subsequent work should also place GraphKAN before the linear layer for even better results. Figure 2 The training time of KSGC and SGC was compared. KSGC's training time increased on Cora, PubMed, and Citeseer. This is due to the additional overhead introduced by GraphKAN's Fourier kernel calculations, but it is still within an acceptable range. Further optimization and improvement can be achieved through other methods such as sparsification and parallelization in the future.

[0114]

[0115] according to Figure 4 The experimental results lead to the following conclusion: The placement of the KAN and linear layers affects node classification accuracy. Adding a linear layer before the KAN layer did not improve model performance; instead, it may have affected the model's effectiveness. The figure compares the accuracy changes across different datasets with different placements of the linear layer and KAN. For the PubMed, Citeseer, and Cora datasets, there are significant differences in accuracy before and after the linear layer is added. Before the addition of the linear layer, KAN achieved low accuracy across all datasets, with a particularly significant difference on the PubMed dataset. However, after the addition of the linear layer, KAN showed significant improvements, particularly on the Cora dataset. This suggests that placing the KAN before the linear layer is more effective in improving accuracy. Therefore, in this model design, placing the KAN layer before the linear layer may be more beneficial for improving model performance, leaving more room for exploration in other models.

[0116] according to Figure 5The experimental results indicate that the computational efficiency of the KAN model is significantly lower than that of the original model. The figure shows the difference in training time between the KSGC model and the SGC model on different datasets. For the PubMed, Cora, and Citeseer datasets, the training time of the KSGC model (red and green bars) is significantly higher than that of the SGC model (blue and purple bars). In particular, the training time of the KSGC model is significantly higher than that of the SGC model on the PubMed and Cora datasets, indicating a time disadvantage for the KSGC model compared to the SGC model. Both the training time and total training time of the KSGC model (green and blue bars) are higher than those of the SGC model, indicating a relatively more complex processing. This may be due to the increased computational effort of the KAN layer, as experiments have shown that KAN is computationally less efficient than some baseline models in some cases. Although the KAN layer enhances model performance and feature extraction capabilities, its computational efficiency still needs to be improved compared to traditional models such as MLP. Therefore, although the KAN model shows strong performance in some tasks, there is still a certain gap in efficiency compared with the traditional SGC model and MLP model. There is a lot of room for development in the future, and its efficiency can be improved through further optimization.

[0117] The present invention also provides a data classification device based on linear feature enhancement, such as Figure 3 As shown, the device includes:

[0118] A propagation module 10 is configured to perform linear propagation on the node feature and the corresponding neighbor node feature to obtain a first linear feature and a second linear feature; the node corresponding to the node feature is the first node;

[0119] A mapping module 20, configured to perform nonlinear mapping on the first linear feature and the second linear feature respectively through KAN nonlinear mapping to obtain a first mapping node feature and a second mapping node feature;

[0120] a combining module 30, configured to determine an embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through a graph convolutional network;

[0121] a superposition module 40 for superimposing the embeddings of each layer to obtain a target node feature representation of the first node;

[0122] Determine module 50, for based on Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node;

[0123] The acquisition module 60 is configured to convert the enhanced features through a softmax function to obtain a classification result.

[0124] The present invention further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0125] S10: performing linear propagation on the node feature and the corresponding neighbor node feature to obtain a first linear feature and a second linear feature; the node corresponding to the node feature is the first node;

[0126] S20: performing nonlinear mapping on the first linear feature and the second linear feature respectively through KAN nonlinear mapping to obtain a first mapping node feature and a second mapping node feature;

[0127] S30: Determine the embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through a graph convolutional network;

[0128] S40: superimposing the embeddings of each layer to obtain a target node feature representation of the first node;

[0129] S50: Based on Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node;

[0130] S60: Convert the enhanced features through a softmax function to obtain a classification result.

[0131] A computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0132] S10: performing linear propagation on the node feature and the corresponding neighbor node feature to obtain a first linear feature and a second linear feature; the node corresponding to the node feature is the first node;

[0133] S20: performing nonlinear mapping on the first linear feature and the second linear feature respectively through KAN nonlinear mapping to obtain a first mapping node feature and a second mapping node feature;

[0134] S30: Determine the embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through a graph convolutional network;

[0135] S40: superimposing the embeddings of each layer to obtain a target node feature representation of the first node;

[0136] S50: Based on Sigmoid activation function and The activation function and the target node feature representation determine the enhanced features of the first node;

[0137] S60: Convert the enhanced features through a softmax function to obtain a classification result.

[0138] Figure 6 FIG1 shows an internal structure diagram of a computer device in an embodiment. The computer device can be a terminal or a server. Figure 6 As shown, the computer device includes a processor, a memory and a network interface connected via a system bus. Among them, the memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When the computer program is executed by the processor, the processor can implement a data classification method based on linear feature enhancement. The internal memory may also store a computer program. When the computer program is executed by the processor, the processor can implement a data classification method based on linear feature enhancement. Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0139] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0140] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0141] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A data classification method based on linear feature enhancement is used to classify graph data. The graph data comes from the Pubmed dataset. The graph data contains multiple nodes, each node has a neighbor node, each node corresponds to a node feature, and each neighbor node corresponds to a neighbor node feature. The method is characterized in that: The method comprises: Performing linear propagation on the node feature and the corresponding neighbor node feature to obtain a first linear feature and a second linear feature; the node corresponding to the node feature is the first node; pass KAN Nonlinear mapping performs nonlinear mapping on the first linear feature and the second linear feature respectively to obtain a first mapping node feature and a second mapping node feature; Determining an embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through a graph convolutional network; Superimposing the embeddings of each layer to obtain a target node feature representation of the first node; based on Sigmoid Activation function and The activation function and the target node feature representation determine the enhanced features of the first node; pass softmax The function transforms the enhanced features to obtain a classification result.

2. The data classification method based on linear feature enhancement according to claim 1, characterized in that: The linear propagation of the node feature and the corresponding neighbor node feature to obtain the first linear feature and the second linear feature is achieved by the following expression: in, is the node feature, is the neighbor node feature, is the first linear feature, is the second linear feature, is the adjacency matrix, is the power of the adjacency matrix.

3. The data classification method based on linear feature enhancement according to claim 2, characterized in that: The first mapping node feature and the second mapping node feature obtained by nonlinearly mapping the first linear feature and the second linear feature respectively through KAN nonlinear mapping are realized by the following expressions: in, is the first mapping node feature, is the second mapping node feature, and is the trainable coefficient, is the first linear feature, is the second linear feature, is the grid number of the Fourier basis, is the frequency index.

4. The data classification method based on linear feature enhancement according to claim 1, characterized in that: The determination of the embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through the graph convolutional network is implemented by the following expression: in, Represented as a node In the Layer embedding, is the first mapping node feature, is the second mapping node feature, Gather for neighbors, To transfer node base KAN activation function, is element-wise multiplication.

5. The data classification method based on linear feature enhancement according to claim 4, characterized in that: The superposition of the embeddings of each layer to obtain the target node feature representation of the first node is achieved by the following expression: in, The first node u The target node feature representation is 、 、...、 Represents the first node u Features at different layers, is the total number of layers, Layer is one of the layers.

6. The data classification method based on linear feature enhancement according to claim 5, characterized in that: The based Sigmoid Activation function and The activation function and the target node feature representation determine the enhanced features of the first node through the following expression: = in, is the weighted feature, is the first mapping node feature, is the adaptive weight, for Sigmoid activation function, is the first linear transformation, is the second linear transformation, is the activation function, is the third mapping node feature represented by the target node feature, The first node u The target node feature representation is and is the trainable coefficient, is the grid number of the Fourier basis, is the frequency index.

7. The data classification method based on linear feature enhancement according to claim 6, characterized in that: Said through softmax The function converts the enhanced features to obtain the classification results through the following expression: in, is the classification result, is the weight matrix, The third mapping node feature represented by the target node feature.

8. A data classification device based on linear feature enhancement, for classifying graph data, wherein the graph data comes from a Pubmed dataset, the graph data contains multiple nodes, each node has a neighbor node, each node corresponds to a node feature, and each neighbor node corresponds to a neighbor node feature, characterized in that: The device comprises: a propagation module, configured to perform linear propagation on the node feature and the corresponding neighbor node feature respectively to obtain a first linear feature and a second linear feature; the node corresponding to the node feature is the first node; Mapping module, used to KAN Nonlinear mapping performs nonlinear mapping on the first linear feature and the second linear feature respectively to obtain a first mapping node feature and a second mapping node feature; a combining module, configured to determine an embedding of the first node in each layer of the graph convolutional network by combining the first mapping node feature and the second mapping node feature through a graph convolutional network; a superposition module, configured to superimpose the embeddings of each layer to obtain a target node feature representation of the first node; Determine the module for Sigmoid Activation function and The activation function and the target node feature representation determine the enhanced features of the first node; Get the module for softmax The function transforms the enhanced features to obtain a classification result.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that A computer program is stored, and when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

Images