Graph data feature extraction device and method based on graph contrast learning, computer equipment and storage medium
Through graph comparison learning and data augmentation technology, the problems of high computational complexity and information loss in large-scale graph data processing are solved, and more accurate and efficient graph data feature extraction is achieved, which enhances the generalization of the model ability.
Patent Information
- Application Number
- CN202411368297.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-29
- Publication Date
- 2025-01-14
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional graph neural networks are too complex in processing large-scale graph data and have serious information losses, making it difficult to effectively extract the features of graph data.
Using a method based on graph comparison learning, the original graph data is decomposed and feature masked through the data enhancement module to generate new graph data, and the GNN encoder with shared weights is used for comparison learning, extract the graph data features, and construct The overall goal loss function is used to optimize model parameters.
It improves the accuracy of graph data feature extraction, reduces computational complexity, reduces information loss, and enhances the generalization ability of the model.
Smart Images
Figure CN119312078A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of feature extraction technology, and in particular to a graph data feature extraction device, method, computer equipment and storage medium based on graph contrast learning. Background Art
[0002] Graph data processing constitutes a key research direction in computer science and engineering, and its application range is wide, covering multiple fields such as machine learning and artificial intelligence. In this field, graph data feature extraction is particularly critical, and its purpose is to extract meaningful features from complex graph structure data for subsequent analysis and processing. Graph data is usually presented in the form of a collection of nodes and edges, and its application areas include but are not limited to social networks, bioinformatics networks, and transportation networks.
[0003] In recent years, with the rapid development of Graph Neural Networks (GNNs), graph data feature extraction technology has achieved significant breakthroughs. GNNs can efficiently capture local and global features in graphs by utilizing graph structure information and relying on the propagation mechanism of nodes and edges. However, traditional GNNs still face the problems of high computational complexity and information loss when processing large-scale graph data. Summary of the invention
[0004] The present invention aims to at least solve the technical problem of high computational complexity in the prior art, and in particular innovatively proposes a graph data feature extraction device, method, computer equipment and storage medium based on graph comparative learning.
[0005] In order to achieve the above-mentioned object of the present invention, the present invention provides a graph data feature extraction device based on graph contrast learning, the device comprising:
[0006] A receiving module, used for receiving original image data, the original image data including an adjacency matrix and a feature matrix;
[0007] A data enhancement module, connected to the receiving module, for performing low-rank singular value decomposition and feature masking on the adjacency matrix and the feature matrix respectively, to obtain new graph data, wherein the new graph data includes a target adjacency matrix and a target feature matrix;
[0008] A feature extraction module is connected to the data enhancement module and the receiving module, and is used to input the adjacency matrix, feature matrix, target adjacency matrix and target feature matrix into the GNN encoder with shared weights, respectively, to obtain the feature representation of the original image data and the feature representation of the new image data, to compare and learn the feature representation of the original image data and the feature representation of the new image data, and to extract the features of the image data.
[0009] As an optional embodiment of the present invention, optionally, the feature extraction module includes:
[0010] An initialization unit, used to initialize extraction parameters involved in the feature extraction process;
[0011] A training unit is used to extract the image data features in the original image data and the new image data based on the extraction parameters.
[0012] As an optional embodiment of the present invention, optionally, the feature extraction module further includes obtaining an original negative sample set according to the original image data;
[0013] Using the target adjacency matrix as a similarity matrix, combining the similarity matrix with the original negative sample set to obtain a new negative sample set;
[0014] Establishing a target loss function based on the new negative sample set;
[0015] An overall target loss function is constructed based on the original image data feature representation and the new image data feature representation.
[0016] As an optional embodiment of the present invention, optionally, the expression of the objective loss function is:
[0017]
[0018] Among them, l(,) represents the target loss function, u i and v i denotes the i-th feature representation generated by the original image data and the new image data, respectively, e denotes the natural base, θ(u i ,v i ) represents u i and v i The cosine similarity of , τ represents the temperature coefficient.
[0019] As an optional embodiment of the present invention, optionally, the expression of the overall objective loss function is:
[0020]
[0021] Among them, L represents the overall target loss function, N represents the number of nodes, i represents the i-th feature representation, l(H1,H2) represents the target loss function of H1 and H2, H1 represents the feature representation of the original graph data, and H2 represents the feature representation of the new graph data.
[0022] As an optional embodiment of the present invention, optionally, the device further includes a graph data module, and the graph data module includes:
[0023] A picture data storage unit, used for storing the original picture data;
[0024] Enhanced data storage unit for storing new graph data;
[0025] The feature storage unit is used to store the graph data features extracted by the feature extraction module.
[0026] In another aspect, the present invention further provides a graph data feature extraction method based on graph contrast learning, the method comprising a graph data feature extraction device based on graph contrast learning;
[0027] The method further comprises:
[0028] S1, initializing parameters of the device, acquiring original image data, and inputting the original image data into the device;
[0029] S2, constructing an overall target loss function based on the new negative sample set, the feature representation of the original image data and the feature representation of the new image data, and performing iterative optimization training on the device using the overall target loss function;
[0030] S3, determine whether the number of iterations reaches the upper limit, if so, extract the graph data features;
[0031] If not, determine whether the overall target loss function has reached convergence. If so, extract the graph data features;
[0032] If not, the overall objective loss function is iteratively optimized using a gradient descent algorithm, and whether the number of iterations reaches the upper limit is re-determined based on the optimized overall objective loss function.
[0033] As another optional embodiment of the present invention, optionally, the parameters include hidden layer dimension, GNN encoder, linear projection head, temperature coefficient, threshold parameter, weight decay parameter, feature masking rate, discarding rate, learning rate, maximum number of iterations and tolerance steps.
[0034] Beneficial effects of the present invention: The present invention generates new graph data by performing low-rank decomposition and feature masking on the original graph data through a data enhancement module, so that the model can learn richer graph structure information, thereby improving the accuracy of feature extraction. The present invention uses a graph contrast learning method to enable the model to effectively utilize the newly generated graph data during the training process, reducing the need for direct processing of large-scale graph data, thereby reducing the computational complexity. In the feature extraction module, by contrast learning the feature representation of the original graph data and the feature representation of the new graph data, the model can better capture the global information in the graph data and reduce information loss. By constructing an overall target loss function and combining the feature representation of the original graph data and the feature representation of the new graph data, the model can better generalize to unseen graph data during the training process, thereby enhancing the generalization ability of the model. The present invention provides a variety of parameter settings, such as hidden layer dimension, GNN encoder, linear projection head, etc., so that users can flexibly adjust model parameters according to specific application scenarios to achieve the best feature extraction effect. By introducing a graph contrast learning mechanism, the present invention not only improves the accuracy of feature extraction, but also effectively reduces computational complexity and information loss, while enhancing the generalization ability of the model.
[0035] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:
[0037] Figure 1 It is a structural schematic diagram of a graph data feature extraction device based on graph contrast learning in Example 1 of the present invention;
[0038] Figure 2 It is a flow chart of the feature extraction of graph data in Examples 1 and 2 of the present invention.
[0039] Figure 3 2 is a schematic diagram showing the structure comparison before and after the implementation of Example 1 of the present invention;
[0040] Figure 4 It is a flowchart of a graph data feature extraction method based on graph comparative learning in Example 2 of the present invention. DETAILED DESCRIPTION
[0041] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.
[0042] Example 1
[0043] like Figure 1 As shown, a graph data feature extraction device based on graph contrast learning, the device includes:
[0044] The receiving module is used to receive the original image data, which includes the adjacency matrix and the feature matrix. In this embodiment, the receiving module is a wireless or wired communication module, which is used to receive instructions and original image data sent by the server.
[0045] A data enhancement module, connected to the receiving module, for performing low-rank singular value decomposition and feature masking on the adjacency matrix and the feature matrix respectively, to obtain new graph data, wherein the new graph data includes a target adjacency matrix and a target feature matrix;
[0046] It should be noted that the data enhancement module includes an adjacency matrix A (the matrix size is D rows, D columns, and the element a in the matrix is i,j represents the data in the i-th row and j-th column of the matrix), the original image data feature matrix X (the matrix size is D rows and F columns, and the element x in the matrix i,k represents the data of the i-th row and k-th column in the matrix), and obtains the result after low-rank singular value decomposition and feature masking, including the left singular vector matrix U q (the matrix size is D rows and q columns, q represents the maximum number of singular values retained after truncating the singular value list), the singular value matrix S q (the matrix size is q rows and q columns), the right singular vector matrix V q (The matrix size is D rows and q columns), the new image feature matrix after feature masking And calculate the adjacency matrix of the reconstructed new graph (The matrix size is D rows and D columns). In the feature masking enhancement process, from the Bernoulli distribution B(1-p F ), generate a mask vector m∈{0,1} F , where p F Represents the probability of feature masking. Finally, the feature matrix of the new image after feature masking is obtained Where ⊙ represents the Hadamard product.
[0047] The feature extraction module is connected to the data enhancement module and the receiving module, and is used to input the adjacency matrix, feature matrix, target adjacency matrix and target feature matrix into the GNN encoder with shared weights, obtain the feature representation of the original image data and the feature representation of the new image data, compare and learn the feature representation of the original image data and the feature representation of the new image data, and extract the image data features. The GNN encoder is each existing GNN neural network, whose input layer input is the number of features, and the output is the hidden layer dimension hid, and the output layer input is the hidden layer dimension hid, and the output is the dimension d.
[0048] like Figure 1As shown, it should be noted that after the receiving module receives the original image data, the data enhancement module will process the original image data. The data enhancement module includes a low-rank decomposition unit and a feature masking unit. The low-rank decomposition unit is responsible for low-rank decomposition of the adjacency matrix of the original image data to extract the main features in the image data. The feature masking unit generates new image data by randomly masking some elements in the feature matrix, thereby increasing the diversity of the data. The feature extraction module is responsible for extracting feature representations from the original image data and the new image data. The module includes a graph neural network (GNN) encoder and a linear projection head. The GNN encoder is used to encode the image data and extract the structural features of the image data. The linear projection head maps the encoded features to a low-dimensional space for subsequent contrastive learning. Contrastive learning is completed by the contrast module, which is responsible for constructing the overall target loss function and using the loss function to iteratively optimize the feature extraction module. The overall target loss function combines the feature representation of the original image data and the feature representation of the new image data. Through contrastive learning, the model can be better generalized to unseen image data. The parameter setting module is responsible for setting and adjusting the model parameters, including the hidden layer dimension, GNN encoder, linear projection head, temperature coefficient, threshold parameter, weight decay parameter, feature masking rate, discarding rate, learning rate, maximum number of iterations and number of tolerance steps. By flexibly adjusting these parameters, users can optimize the model performance according to specific application scenarios. The graph data module includes a graph data storage unit, an enhanced data storage unit and a feature storage unit. The graph data storage unit is used to store the original graph data, the enhanced data storage unit is used to store the new graph data processed by the data enhancement module, and the feature storage unit is used to store the graph data features extracted by the feature extraction module. In the graph data feature extraction method, the parameters are first initialized and the original graph data is obtained. Then, the overall target loss function is constructed based on the new negative sample set, the original graph data feature representation and the new graph data feature representation. The device is iteratively optimized and trained using the loss function until the number of iterations reaches the upper limit or the overall target loss function converges. Finally, the graph data features are extracted. By introducing the graph contrast learning mechanism, this embodiment not only improves the accuracy of feature extraction, but also effectively reduces the computational complexity and information loss, while enhancing the generalization ability of the model. By flexibly adjusting parameter settings, users can optimize model performance according to specific application scenarios to achieve the best feature extraction effect.
[0049] As an optional embodiment of the present invention, optionally, the feature extraction module includes:
[0050] An initialization unit, used to initialize extraction parameters involved in the feature extraction process;
[0051] It should be noted that the extraction parameters involved in the initialization unit include the output dimension d of the graph data feature, the hidden layer dimension hid, the GNN encoder, the linear projection head h, the temperature coefficient τ, the threshold parameter k, the weight decay parameter wd, the feature masking rate mr, the discarding rate dr, the learning rate lr, the maximum number of iterations max, and the tolerance step number patience, where:
[0052] The graph data feature output dimension d is the dimension of the final output graph data feature and is initialized to a positive integer, such as 16.
[0053] The hidden layer dimension hid is the dimension of the graph data feature transformation and is initialized to a positive integer, such as 32.
[0054] The GNN encoder is an existing GNN method, whose input layer input is the number of features and output is the hidden layer dimension hid, and the output layer input is the hidden layer dimension hid and output is the dimension d.
[0055] The linear projection head h is a linear projection layer for contrastive learning after the GNN encoder, with an input of dimension d and an output of dimension d.
[0056] The temperature coefficient τ is a constant that controls the degree of perception of the device to negative samples and is initialized to a positive number, such as 0.5.
[0057] The threshold parameter k is a constant that controls the threshold when selecting negative samples and is initialized to a positive number less than 1, such as 0.5.
[0058] The weight decay parameter wd is a constant used to regularize the device to prevent the device from overfitting, and is initialized to a positive number less than 0.01, such as 0.001.
[0059] The feature masking rate mr is a constant that controls the number of randomly masked features and is initialized to a positive number less than 1, such as 0.3.
[0060] The discard rate dr is a constant that prevents the device from overfitting and enhances the robustness of the device, and is initialized to a positive number less than 1, such as 0.5.
[0061] The learning rate lr is a constant that controls the update step size of the device parameters during the extraction iteration process and is initialized to a positive number less than 0.1, such as 0.05.
[0062] The maximum number of iterations max is a constant for the maximum number of iterations of the control device and is initialized to a positive integer, such as 1000.
[0063] The tolerance step number patience is a constant that controls the device to stop training early and is initialized to a positive integer, such as 100.
[0064] A training unit is used to extract the image data features in the original image data and the new image data based on the extraction parameters.
[0065] It should be noted that the training unit will adopt the optimization strategy of graph contrast learning in the process of extracting graph data features. The core idea of this strategy is to enable the model to better capture the structural information and feature information in the graph data through the contrast learning mechanism. Specifically, the training unit will first generate a negative sample set, which contains graph data with similar structure but different features to the original graph data. Then, the training unit will use positive samples and negative samples to train the model, and optimize the model parameters by minimizing the distance between positive samples and maximizing the distance between positive samples and negative samples. During the training process, the training unit will continuously adjust the extraction parameters to ensure that the model can effectively learn the key features in the graph data. For example, by adjusting the temperature coefficient, the model's perception of negative samples can be controlled, thereby affecting the model's ability to distinguish. At the same time, by adjusting the feature masking rate, the number of features randomly masked by the model during the training process can be controlled to improve the generalization ability of the model. In addition, the training unit will monitor the training process of the model to ensure that the model does not overfit. For example, by adjusting the dropout rate and weight decay parameters, the robustness of the model can be enhanced to prevent the occurrence of overfitting. At the same time, the training unit will control the update speed and training time of the model parameters according to the learning rate and the maximum number of iterations to ensure that the model can achieve the best performance in a limited time. Finally, the training unit will output optimized graph data features, which can not only accurately reflect the structure and feature information of the original graph data, but also have high generalization ability and robustness. In this way, the present invention can effectively solve the key problems in graph data feature extraction and provide strong support for subsequent graph data processing and analysis tasks.
[0066] As an optional embodiment of the present invention, optionally, the feature extraction module further includes obtaining an original negative sample set according to the original image data;
[0067] The expression of the original negative sample set is:
[0068] NSR(u)={v|v≠u, v≠u′, u, v∈V1∪V2}
[0069] Among them, NSR(u) represents the original negative sample set;
[0070] u,v represent nodes in V1 and V2;
[0071] V1 represents the node set in the original graph data;
[0072] V2 represents the node set in the new graph data;
[0073] u′ represents a node in V2 other than the anchor node corresponding to the u node in V1.
[0074] Using the target adjacency matrix as a similarity matrix, combining the similarity matrix with the original negative sample set to obtain a new negative sample set;
[0075] The expression of the new negative sample set is:
[0076] NS(u)={v|S(u,v) max *k,k,v∈NSR(u)}
[0077] Among them, NS(u) represents the new negative sample set;
[0078] u,v represent nodes in V1 and V2;
[0079] V1 represents the node set in the original graph data;
[0080] V2 represents the node set in the new graph data;
[0081] S(u,v) represents the similarity matrix of nodes u and v;
[0082] S max Represents the maximum value of the similarity matrix S;
[0083] k represents the threshold parameter.
[0084] Establishing a target loss function based on the new negative sample set;
[0085] An overall target loss function is constructed based on the original image data feature representation and the new image data feature representation.
[0086] It should be noted that after the overall objective loss function is established, the model is iteratively optimized using the stochastic gradient descent method to maximize the overall loss value. Figure 2 As shown, repeat Figure 2 In the training process, the new sample set is calculated only once in the same training, and the training is continued until convergence. The convergence condition is that the maximum number of iterations max is reached, or before the start of this iteration, the number of consecutive steps without improvement has reached the tolerance step number patience.
[0087] As an optional embodiment of the present invention, optionally, the expression of the objective loss function is:
[0088]
[0089] Among them, l(,) represents the target loss function;
[0090] u i and v i Respectively represent the i-th feature representation generated by the original image data and the new image data;
[0091] e represents the natural base;
[0092] θ(u i ,v i ) represents u i and v i The cosine similarity of
[0093] τ is expressed as the temperature coefficient.
[0094] As an optional embodiment of the present invention, optionally, the expression of the overall objective loss function is:
[0095]
[0096] Where L represents the overall target loss function;
[0097] N represents the number of nodes;
[0098] i represents the i-th feature representation;
[0099] l(H1,H2) represents the target loss function of H1 and H2;
[0100] H1 represents the feature representation of the original image data;
[0101] H2 represents the feature representation of new graph data.
[0102] As an optional embodiment of the present invention, optionally, the device further includes a graph data module, and the graph data module includes:
[0103] A picture data storage unit, used for storing the original picture data;
[0104] Enhanced data storage unit for storing new graph data;
[0105] The feature storage unit is used to store the graph data features extracted by the feature extraction module.
[0106] like Figure 3 As shown, in order to verify the performance of the graph data feature extraction device and method based on graph contrast learning, the staff installed the device on a PC (configuration: INTEL i7-13700, 2.1GHz processor, 32G memory) and ran a simulation experiment for example analysis. In the example analysis, the graph data used is a real graph data that actually exists. The example analysis uses the node classification accuracy ACC as the evaluation index of the accuracy of graph data feature extraction. The higher the ACC, the higher the accuracy of graph data feature extraction.
[0107] Figure 3 The accuracy of feature extraction on real image data is compared before and after applying this embodiment. Figure 3After applying this embodiment, when extracting graph data features, using the original GNN method and applying the original GNN method to this embodiment will result in different accuracy results, resulting in different extraction accuracies. The first case represents the result of extracting graph data features using the original GNN method, and the second case represents the result of extracting graph data features with higher accuracy after implementing the invention.
[0108] It can be seen from the above technical solution that an embodiment of the present invention provides a graph data feature extraction device based on graph contrast learning, which specifically acts on graph data and can perform graph data feature extraction with high accuracy in accordance with statistical laws to solve the problem of graph data feature extraction in complex graph structures.
[0109] Example 2
[0110] like Figure 2 and 4 As shown, a graph data feature extraction method based on graph contrast learning, the method includes a graph data feature extraction device based on graph contrast learning;
[0111] The method further comprises:
[0112] S1, initializing parameters of the device, acquiring original image data, and inputting the original image data into the device;
[0113] It should be noted that the initialized parameters include the output dimension d of the graph data feature, the hidden layer dimension hid, the GNN encoder, the linear projection head h, the temperature coefficient τ, the threshold parameter k, the weight decay parameter wd, the feature masking rate mr, the discarding rate dr, the learning rate lr, the maximum number of iterations max, and the tolerance step number patience.
[0114] S2, constructing an overall target loss function based on the new negative sample set, the feature representation of the original image data and the feature representation of the new image data, and performing iterative optimization training on the device using the overall target loss function;
[0115] S3, determine whether the number of iterations reaches the upper limit, if so, extract the graph data features;
[0116] If not, determine whether the overall target loss function has reached convergence. If so, extract the graph data features;
[0117] If not, the overall objective loss function is iteratively optimized using a gradient descent algorithm, and whether the number of iterations reaches the upper limit is re-determined based on the optimized overall objective loss function.
[0118] like Figure 2 and 4As shown, it should be noted that in step S2, the new negative sample set is generated by randomly selecting nodes from the original graph data and constructing a graph data feature representation that is dissimilar to the original graph data feature representation. The generation of the new negative sample set helps to improve the discrimination of graph data feature extraction, thereby improving the overall feature extraction accuracy.
[0119] In step S3, the upper limit of the number of iterations and the convergence condition are set to ensure the stability and convergence of the training process. The upper limit of the number of iterations can prevent the training process from falling into an infinite loop, while the convergence condition ensures that the model stops training after reaching a certain accuracy to avoid overfitting. After extracting the graph data features, the features are stored in the feature storage unit for subsequent graph data processing and analysis tasks.
[0120] In summary, this embodiment provides a graph data feature extraction method based on graph contrast learning, which can effectively extract high-quality graph data features by constructing an overall target loss function and performing iterative optimization training. This method not only improves the accuracy of graph data feature extraction, but also has good generalization ability, is suitable for a variety of graph data processing tasks, and has broad application prospects.
[0121] As another optional embodiment of the present invention, optionally, the parameters include hidden layer dimension, GNN encoder, linear projection head, temperature coefficient, threshold parameter, weight decay parameter, feature masking rate, discarding rate, learning rate, maximum number of iterations and tolerance steps.
[0122] Example 3
[0123] A computer device comprising:
[0124] processor;
[0125] a memory for storing processor-executable instructions;
[0126] Among them, the processor is configured to implement a graph data feature extraction method based on graph contrast learning in Example 1 when executing the executable instructions.
[0127] It should be noted that the computer device includes: a processor, a memory, and the computer device may also include one or more of a multimedia component, an input / output (I / O) interface, and a communication component.
[0128] Among them, the processor is used to control the overall operation of the computer device to complete all or part of the steps in the above-mentioned big data-based factory equipment automated testing method.
[0129] The memory is used to store various types of data to support operations on the computer device, which data may include, for example, instructions for any application or method operating on the computer device, as well as application-related data; the memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.
[0130] The multimedia component may include a screen and an audio component, wherein the screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals; for example, the audio component may include a microphone for receiving external audio signals, and the received audio signals may be further stored in a memory or sent via a communication component; the audio component also includes at least one speaker for outputting audio signals.
[0131] The I / O interface provides an interface between the processor and other interface modules, which may be a keyboard, a mouse, buttons, etc.; these buttons may be virtual buttons or physical buttons.
[0132] The communication component is used for wired or wireless communication between the computer device and other devices; wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G or 5G, or a combination of one or more of them, so the corresponding communication component may include: Wi-Fi module, Bluetooth module, NFC module, mobile phone communication module.
[0133] As a preferred solution of this embodiment, the computer device can be implemented by one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), controllers, microcontrollers, microprocessors or other electronic components to execute the above-mentioned big data-based factory equipment automation testing method.
[0134] Example 4
[0135] A computer-readable storage medium, comprising:
[0136] a memory having a computer program stored thereon;
[0137] A processor is used to execute the program in the memory to implement a graph data feature extraction method based on graph contrast learning in Example 1.
[0138] It should be noted that the electronic device according to the embodiment of the present disclosure includes a processor and a memory for storing processor executable instructions, wherein the processor is configured to implement any of the above-mentioned cloud resource task scheduling methods with efficient optimization when executing the executable instructions.
[0139] Here, it should be noted that the number of processors can be one or more. At the same time, the electronic device in the embodiment of the present disclosure may also include an input device and an output device. Among them, the processor, memory, input device and output device may be connected through a bus or in other ways, which are not specifically limited here.
[0140] As a computer-readable storage medium, the memory can be used to store software programs, computer executable programs and various modules, such as: the program or module corresponding to the cloud resource task scheduling method with efficient optimization in the embodiment of the present disclosure. The processor executes various functional applications and data processing of the electronic device by running the software programs or modules stored in the memory.
[0141] The input device can be used to receive input numbers or signals. The signal can be a key signal related to user settings and function control of the device / terminal / server. The output device can include a display device such as a display screen.
[0142] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.
Claims
1. A graph data feature extraction device based on graph contrast learning, characterized in that: The device comprises: A receiving module, used for receiving original image data, the original image data including an adjacency matrix and a feature matrix; A data enhancement module, connected to the receiving module, for performing low-rank singular value decomposition and feature masking on the adjacency matrix and the feature matrix respectively, to obtain new graph data, wherein the new graph data includes a target adjacency matrix and a target feature matrix; A feature extraction module is connected to the data enhancement module and the receiving module, and is used to input the adjacency matrix, feature matrix, target adjacency matrix and target feature matrix into the GNN encoder with shared weights, respectively, to obtain the feature representation of the original image data and the feature representation of the new image data, to compare and learn the feature representation of the original image data and the feature representation of the new image data, and to extract the features of the image data.
2. The graph data feature extraction device based on graph contrast learning according to claim 1, characterized in that: The feature extraction module comprises: An initialization unit, used to initialize extraction parameters involved in the feature extraction process; A training unit is used to extract the image data features in the original image data and the new image data based on the extraction parameters.
3. A graph data feature extraction device based on graph contrast learning as claimed in claim 1 or 2, characterized in that: The feature extraction module further comprises obtaining an original negative sample set according to the original image data; Using the target adjacency matrix as a similarity matrix, combining the similarity matrix with the original negative sample set to obtain a new negative sample set; Establishing a target loss function based on the new negative sample set; An overall target loss function is constructed based on the original image data feature representation and the new image data feature representation.
4. The graph data feature extraction device based on graph contrast learning according to claim 3, characterized in that: The expression of the objective loss function is: Among them, l(,) represents the target loss function, u i and v i denotes the i-th feature representation generated by the original image data and the new image data, respectively, e denotes the natural base, θ(u i ,v i ) represents u i and v i The cosine similarity of , τ represents the temperature coefficient.
5. The graph data feature extraction device based on graph contrast learning according to claim 3, characterized in that: The expression of the overall objective loss function is: Among them, L represents the overall target loss function, N represents the number of nodes, i represents the i-th feature representation, l(H1,H2) represents the target loss function of H1 and H2, H1 represents the feature representation of the original graph data, and H2 represents the feature representation of the new graph data.
6. The graph data feature extraction device based on graph contrast learning according to claim 1, characterized in that: The device further includes a graph data module, wherein the graph data module includes: A picture data storage unit, used for storing the original picture data; Enhanced data storage unit for storing new graph data; The feature storage unit is used to store the graph data features extracted by the feature extraction module.
7. A graph data feature extraction method based on graph contrast learning, characterized in that: The method comprises a graph data feature extraction device based on graph contrast learning as described in any one of claims 1 to 6; The method further comprises: S1, initializing parameters of the device, acquiring original image data, and inputting the original image data into the device; S2, constructing an overall target loss function based on the new negative sample set, the feature representation of the original image data and the feature representation of the new image data, and performing iterative optimization training on the device using the overall target loss function; S3, determine whether the number of iterations reaches the upper limit, if so, extract the graph data features; If not, determine whether the overall target loss function has reached convergence. If so, extract the graph data features; If not, the overall objective loss function is iteratively optimized using a gradient descent algorithm, and whether the number of iterations reaches the upper limit is re-determined based on the optimized overall objective loss function.
8. The method for extracting graph data features based on graph contrast learning according to claim 7, characterized in that: The parameters include hidden layer dimension, GNN encoder, linear projection head, temperature coefficient, threshold parameter, weight decay parameter, feature masking rate, dropout rate, learning rate, maximum number of iterations and tolerance steps.
9. A computer device, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to implement a graph data feature extraction method based on graph contrast learning as described in any one of claims 7 to 8 when executing the executable instructions.
10. A computer-readable storage medium, characterized in that: include: a memory having a computer program stored thereon; A processor, used to execute the program in the memory to implement a graph data feature extraction method based on graph contrast learning as described in any one of claims 7 to 8.