Dynamic graph data feature extraction method, device and equipment based on lightweight dynamic graph neural network

Abstract

The invention provides a dynamic graph data feature extraction method, device and equipment based on a lightweight dynamic graph neural network. The method comprises the following steps: converting dynamic data into dynamic graph structure data; constructing a loss function; performing convolution operation on the dynamic graph structure data to extract dynamic graph data features; feature transformation operation and nonlinear activation operation in convolution operation are removed, and dynamic graph data features are extracted; judging whether the loss value of the loss function is converged or not or whether the number of iterations reaches an upper line or not; if yes, outputting dynamic graph data features; and if not, carrying out iterative optimization on the loss function by utilizing a stochastic gradient descent method. According to the method, by reducing the feature transformation operation and the nonlinear activation operation, the calculation complexity is remarkably reduced, so that the calculation complexity is reduced, and the training efficiency of the model is improved. In addition, by introducing the symmetrically standardized adjacent tensor, the description capability of the model on the cross-space-time association relationship in the dynamic graph data is enhanced, so that the extracted features are more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic graph data feature extraction, and in particular to a dynamic graph data feature extraction method, device and equipment based on a lightweight dynamic graph neural network. Background Art

[0002] In today's information age, the generation and transmission of data present dynamic characteristics, that is, data is constantly changing in time and space. Especially in modern large-scale complex systems, such as social networks, transportation systems, the Internet of Things, etc., the data in these systems are usually dynamic and complex. Efficient and accurate feature extraction of these dynamic data is of great significance for analyzing and understanding system behavior, predicting future trends, and making intelligent decisions. Since this type of data can be converted into graph structured data and graph neural networks have become a hot topic of research due to their advantages in processing graph structured data. Graph neural networks can effectively capture the relationship between nodes and the topological structure information of the graph, thus achieving remarkable results in the analysis of static graph data. However, the characteristics of dynamic graph data (i.e., dynamic changes in time and space) put higher requirements on existing graph neural networks.

[0003] Traditional graph neural network models mainly process static graph data and cannot make full use of information in the time dimension. In dynamic graphs, the attributes of nodes and edges change over time, which requires the model to not only capture spatial relationships but also be able to handle dynamic changes in time. Therefore, for feature extraction of dynamic graph data, it is necessary to develop dynamic graph neural network models that can effectively model in the spatiotemporal dimension. At present, some studies have attempted to expand traditional graph neural networks to adapt to dynamic graph data, such as dynamic graph convolutional networks and time-series graph neural networks. These models attempt to capture the time-varying characteristics of nodes in dynamic graph data by introducing mechanisms such as timestamps and dynamic adjacency matrices. However, these methods only consider the propagation of spatial information on a single time slice, hindering the direct information flow across time and space, and cannot accurately describe complex cross-time and space associations. In addition, these methods still have problems such as high computational complexity and long model training time when processing large-scale data, which limits their wide application in practical applications. 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 particularly innovatively proposes a method, device and equipment for extracting dynamic graph data features based on a lightweight dynamic graph neural network.

[0005] In order to achieve the above-mentioned object of the present invention, the present invention provides a method for extracting dynamic graph data features based on a lightweight dynamic graph neural network, the method comprising:

[0006] S1, collecting dynamic data, and converting the dynamic data into dynamic graph structure data;

[0007] S2. Constructing a loss function based on the dynamic graph structure data;

[0008] S3, performing a convolution operation on the dynamic graph structure data to extract node features;

[0009] S4, removing the feature transformation operation and the nonlinear activation operation in the convolution operation to extract dynamic graph data features;

[0010] S5. Determine whether the loss value of the loss function converges or whether the number of iterations reaches an upper limit;

[0011] If so, output the dynamic graph data features;

[0012] If not, the loss function is iteratively optimized using stochastic gradient descent.

[0013] As an optional embodiment of the present invention, optionally, the expression of the loss function is:

[0014]

[0015] Among them, ω represents the loss value, i represents the i-th time slice graph in the dynamic graph structure data, and Y i represents the true weight label of the i-th time slice graph in the dynamic graph structure data, log() represents the logarithmic function, represents the predicted weight label of the i-th time slice graph in the dynamic graph structure data, λ represents the regularization factor, || || F represents the F norm, H 0 Represents the initialized dynamic graph feature tensor.

[0016] As an optional embodiment of the present invention, optionally, the expression of the convolution operation in step S3 is:

[0017]

[0018] Among them, H l+1 represents the output feature of the l+1th layer convolution operation, σ() represents the nonlinear activation function operation, Represents the symmetric normalized adjacency tensor used to perform the l-th layer of convolution, * represents the tensor multiplication operator, H l represents the output feature of the l-th layer convolution operation, W l represents the feature transformation tensor of the lth layer, ×3 represents M transformation operations, M represents the temporal transformation matrix, and · represents the area operation.

[0019] As an optional embodiment of the present invention, optionally, the method further includes symmetric normalization of the adjacency tensor of the dynamic graph structure data; the expression of the symmetric normalization is:

[0020]

[0021] in, represents the adjacency tensor after symmetric normalization, D represents the degree tensor constructed based on the adjacency tensor, · represents the area operation, and A represents the adjacency tensor of the dynamic graph structure data.

[0022] As an optional embodiment of the present invention, optionally, in step S4, the expression for removing the feature transformation operation and the nonlinear activation operation in the convolution operation is:

[0023]

[0024] Among them, H l+1 represents the output features of the l+1th layer convolution operation, Represents the symmetric normalized adjacency tensor used to perform the l-th layer of convolution, * represents the tensor multiplication operator, H l represents the output features of the l-th convolution operation, ×3 represents M conversion operations, M represents the time conversion matrix, and · represents the area operation.

[0025] As an optional embodiment of the present invention, optionally, the expression of the M conversion operation is:

[0026]

[0027] in, represents the output of the l+1th layer M conversion operation, t represents the tth slice of the element of the required M conversion operation in the time dimension, k represents the index of the M conversion operation from the first slice to the tth slice in the time dimension, and m tk represents the value of k at position t in the time conversion matrix, Represents the input of the l-th layer M transformation operation.

[0028] On the other hand, the present invention also provides a dynamic graph data feature extraction device based on a lightweight dynamic graph neural network, the device comprising a dynamic graph data feature extraction method based on a lightweight dynamic graph neural network; the device also comprises:

[0029] A data receiving module, used for acquiring dynamic data and converting the dynamic data into dynamic graph structure data;

[0030] An objective function module, connected to the data receiving module, for constructing a loss function using the dynamic graph structure data;

[0031] A feature extraction module, connected to the data receiving module and the objective function module, for performing a convolution operation on the dynamic graph structure data to extract dynamic graph data features;

[0032] The storage module is connected to the data receiving module and the data receiving module, and is used to store the dynamic data, dynamic graph structure data and dynamic graph data features.

[0033] As another optional embodiment of the present invention, optionally, the feature extraction module includes:

[0034] A parameter initialization unit, used to initialize the relevant parameters for extracting dynamic graph data features;

[0035] A lightweight dynamic graph neural network unit, connected to the parameter initialization unit, for extracting the dynamic graph data features based on the dynamic graph structure data and the initialized related parameters;

[0036] The output unit is connected to the lightweight dynamic graph neural network unit and is used to extract the most accurate dynamic graph data features.

[0037] As another optional embodiment of the present invention, optionally, the storage module includes:

[0038] A dynamic graph data storage unit, used for storing the dynamic graph structure data;

[0039] The feature storage unit is used to store the dynamic graph data features.

[0040] The beneficial effect of the present invention is that the present invention significantly reduces the computational complexity by reducing feature transformation operations and nonlinear activation operations, thereby reducing the computational complexity and improving the training efficiency of the model. In addition, by introducing symmetrically standardized adjacency tensors, the model's ability to describe cross-temporal and spatial correlation relationships in dynamic graph data is enhanced, making the extracted features more accurate. The device and method of the present invention are not only suitable for feature extraction of large-scale dynamic graph data, but also have broad application prospects in practical applications, such as in the fields of social network analysis, traffic flow prediction, bioinformatics, etc.

[0041] 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

[0042] 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:

[0043] Figure 1It is a flow chart of a method for extracting dynamic graph data features based on a lightweight dynamic graph neural network according to embodiment 1 of the present invention.

[0044] Figure 2 This is another flow chart of the method for extracting dynamic graph data features based on a lightweight dynamic graph neural network according to embodiment 1 of the present invention.

[0045] Figure 3 It is a structural schematic diagram of a dynamic graph data feature extraction method based on a lightweight dynamic graph neural network according to embodiment 2 of the present invention. DETAILED DESCRIPTION

[0046] 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.

[0047] Example 1

[0048] like Figure 1 As shown, a dynamic graph data feature extraction method based on a lightweight dynamic graph neural network, the method comprising:

[0049] S1, collecting dynamic data, and converting the dynamic data into dynamic graph structure data;

[0050] It should be noted that the dynamic data collected in step S1 can be of various types, such as user behavior data in social networks, vehicle flow data in traffic networks, etc. These data usually contain rich spatiotemporal information and can reflect the characteristics of dynamic changes. Specifically, the data is initialized into a dynamic graph structure, that is, the data is initialized into an adjacency tensor A, whose shape is N×N×T (slice A in the tensor t represents the adjacency matrix of the graph at time t, A t The element a in the matrix ijt Indicates the connection between node i and node j in the graph at time t. If node i and node j are connected in the graph, a ijt Equal to 1; otherwise, if node i and node j are not connected in the graph, a ijt equal to 0).

[0051] S2. Constructing a loss function based on the dynamic graph structure data;

[0052] It should be noted that in step S2, based on the input dynamic graph data, this embodiment mainly uses the weight prediction accuracy of the edges between nodes in the dynamic graph to evaluate and guide the dynamic graph data feature extraction process. For a target dynamic graph, let X iRepresents the learned dynamic graph data features, where i represents the i-th dynamic graph and its true weight label is Y i , the dynamic graph data features learned from the input dynamic graph data can be passed through a predictor (such as a fully connected network) to obtain the predicted dynamic graph weight label Then the cross entropy loss function is used for optimization, which is expressed as follows:

[0053]

[0054] Among them, DisFunction represents the true weight Y i With prediction weight The distance function between is used to calculate the accuracy between the predicted weights and the true weights.

[0055] S3, performing a convolution operation on the dynamic graph structure data to extract node features;

[0056] It should be noted that in step S3, a convolution operation is performed on the dynamic graph structure data to extract node features. This embodiment adopts a lightweight dynamic graph convolution network structure, which can effectively process the spatiotemporal information in the dynamic graph data. Specifically, the structure extends the traditional two-dimensional convolution operation to three-dimensional space by introducing the time dimension, so as to capture the feature changes of nodes in different time slices.

[0057] S4, removing the feature transformation operation and the nonlinear activation operation in the convolution operation to extract dynamic graph data features;

[0058] It should be noted that in step S4, this embodiment further simplifies the model structure by removing the feature transformation operation and nonlinear activation operation in the convolution operation, thereby realizing the rapid extraction of dynamic graph data features. This lightweight dynamic graph convolution network structure not only reduces the computational complexity of the model, but also improves the efficiency of feature extraction. In this way, this embodiment can quickly extract useful feature information from dynamic graph data, providing a solid foundation for subsequent analysis and processing.

[0059] S5. Determine whether the loss value of the loss function converges or whether the number of iterations reaches an upper limit;

[0060] If so, output the dynamic graph data features;

[0061] If not, the loss function is iteratively optimized using stochastic gradient descent.

[0062] like Figure 2As shown, it should be noted that in step S5, this embodiment uses a lightweight convolution operation to extract node features. The spatial dependencies of nodes in different time slices can be effectively captured by using the symmetric normalization of the adjacency tensor and the M conversion operation. Specifically, the adjacency tensor is first symmetrically normalized to ensure that the importance of each node is reasonably reflected. Then, through the M conversion operation, information supplementation in the time domain is achieved, so that the graph structure in each time slice can perceive each other in the time dimension, thereby improving the accuracy of feature extraction.

[0063] After completing the above steps, the obtained dynamic graph adjacency tensor and feature tensor are used as input to perform a lightweight dynamic graph convolution operation. This operation captures the spatial dependencies of nodes through the tensor multiplication operator, combined with the feature transformation tensor, and enhances the expressiveness of the model through a nonlinear activation function. However, in order to further reduce the computational complexity and optimize memory usage efficiency, the lightweight dynamic graph convolution module of the present invention removes the feature transformation and nonlinear activation operations to obtain a more lightweight model.

[0064] During the model training process, this embodiment controls the training process by determining whether the current number of training rounds of the model has reached the upper limit, or whether the model parameters (loss values) have converged on the dynamic graph data set. When the model parameters converge, a gradient descent method such as Adam is used to iteratively optimize to obtain the most accurate dynamic graph feature results. Finally, the extracted dynamic graph feature results are stored to provide efficient and accurate data support for subsequent dynamic graph analysis and processing.

[0065] like Figure 1 and 2 As shown, in summary, this embodiment significantly reduces the computational complexity by reducing feature transformation operations and nonlinear activation operations, thereby reducing the computational complexity and improving the training efficiency of the model. In addition, by introducing symmetrically standardized adjacency tensors, the model's ability to describe cross-temporal and spatial correlation relationships in dynamic graph data is enhanced, making the extracted features more accurate. The device and method of the present invention are not only suitable for feature extraction of large-scale dynamic graph data, but also have broad application prospects in practical applications, such as in the fields of social network analysis, traffic flow prediction, bioinformatics, etc.

[0066] As an optional embodiment of the present invention, optionally, the expression of the loss function is:

[0067]

[0068] Among them, ω represents the loss value;

[0069] i represents the i-th time slice graph in the dynamic graph structure data;

[0070] Y i Represents the true weight label of the i-th time slice graph in the dynamic graph structure data;

[0071] log() represents the logarithmic function;

[0072] Represents the predicted weight label of the i-th time slice graph in the dynamic graph structure data;

[0073] λ represents the regularization factor;

[0074] |||| F represents the F-norm;

[0075] H 0 Represents the initialized dynamic graph feature tensor.

[0076] It should be noted that L2 regularization is used in the above formula to constrain the optimization process to avoid overfitting of the model.

[0077] As an optional embodiment of the present invention, optionally, the expression of the convolution operation in step S3 is:

[0078]

[0079] Among them, H l+1 Represents the output features of the l+1th layer convolution operation;

[0080] σ() represents the nonlinear activation function operation;

[0081] Represents the symmetric normalized adjacency tensor used when performing the l-th layer of convolution;

[0082] * represents the tensor multiplication operator;

[0083] H l Represents the input features of the l-th layer convolution operation;

[0084] W l Represents the feature transformation tensor of layer l;

[0085] ×3 indicates M conversion operation;

[0086] M represents the time transformation matrix;

[0087] × represents area operation.

[0088] It should be noted that before the convolution operation, the M conversion operation is used to obtain the dynamic graph adjacency tensor after the time domain information is supplemented, that is, the adjacency tensor of the next layer is expressed as:

[0089]

[0090] in, Represents the output adjacency tensor of the l-th layer M transformation operation;

[0091] A l Represents the adjacency tensor used when performing the l-th layer of lightweight dynamic graph convolution;

[0092] It means that the initial symmetric normalized adjacency tensor M is the time transformation matrix. Specifically, the specific calculation operation of the above transformation process is:

[0093]

[0094] in, represents the output of the M-th transformation operation at layer l;

[0095] t represents the t-th slice of the element in the time dimension that requires the M transformation operation;

[0096] k represents the index of the M transformation operation from the first slice to the tth slice of the time dimension;

[0097] m tk represents the value of the time transformation matrix at position t, k;

[0098] Represents the input of the l-th layer M transformation operation.

[0099] The above M conversion operation can realize the mutual perception of the topological structure of the graph in each time slice in the time domain, thereby improving the accuracy of dynamic graph features. Similarly, in order to ensure that the graph features between time slices can communicate seamlessly, each layer of dynamic graph features will also perform the following M conversion:

[0100] H l+1 =H l ×3M

[0101] Among them, H l+1 represents the output of the M-th transformation operation at the l+1th layer;

[0102] H l represents the input of the l-th layer M conversion operation;

[0103] According to the above formula, the elements in each layer of dynamic graph features will be accurately calculated as:

[0104]

[0105] According to the above M conversion operation, we can obtain the dynamic graph adjacency tensor and dynamic graph feature tensor after M conversion. Then, these tensors are used as input to perform dynamic graph convolution operations to capture the spatial dependency relationship of nodes in the graph. The specific calculation process is as follows:

[0106] H l+1 =A l+1 ·H l+1

[0107] In this way, each node in a single time slice graph can capture the characteristics of other nodes in a similar area in terms of spatiotemporal position or some nodes in other time slices. The specific calculation process of the dynamic graph convolution operation is as follows:

[0108]

[0109] in, Represents the features of node i in the fth feature dimension on the tth time slice obtained by the lth layer dynamic graph convolution operation;

[0110] u represents the neighbor node u of node i in the l+1th layer at the tth time slice after M conversion;

[0111] N(i,l+1,t) represents the number of all neighbors of node i in the l+1th layer at the tth time slice after M conversion;

[0112] Represents the l+1th layer adjacency tensor obtained after M conversion The value between node i and its neighbor node u at the tth time slice;

[0113] Then summarize the above process to get the complete dynamic graph convolution operation based on the tensor multiplication framework, that is, the expression:

[0114] H l+1 =σ(A l *H l *W l )=σ(((A l ×3M)·(H l ×3M))×3(W l ×3M))

[0115] As an optional embodiment of the present invention, optionally, the method further includes symmetric normalization of the adjacency tensor of the dynamic graph structure data; the expression of the symmetric normalization is:

[0116]

[0117] in, represents the adjacency tensor after symmetric normalization;

[0118] D represents the degree tensor constructed based on the adjacency tensor;

[0119] ·Represent area operations;

[0120] A represents the adjacency tensor of dynamic graph structure data.

[0121] It should be noted that in order to avoid the numerical instability and other problems that may arise during the training process when directly using the adjacency tensor to describe complex non-Euclidean topological structure information, the adjacency tensor is first symmetrically standardized, and then a new dynamic graph convolution module is implemented based on the tensor multiplication framework. Finally, in order to reduce the computational complexity of the dynamic graph convolution module and optimize memory usage efficiency, the feature transformation and nonlinear activation operations in the module are removed to obtain the final lightweight dynamic graph convolution module.

[0122] As an optional embodiment of the present invention, optionally, in step S4, the expression for removing the feature transformation operation and the nonlinear activation operation in the convolution operation is:

[0123]

[0124] Among them, H l+1 Represents the output features of the l+1th layer of lightweight dynamic graph convolution;

[0125] Represents the symmetric normalized adjacency tensor used when performing the l-th layer of lightweight dynamic graph convolution;

[0126] * represents the tensor multiplication operator;

[0127] H l Represents the input features of the lightweight dynamic graph convolution at layer l;

[0128] ×3 indicates M conversion operation;

[0129] M represents the time transformation matrix;

[0130] Represents area operations.

[0131] It should be noted that the above-mentioned lightweight dynamic graph convolution method extends the traditional graph convolution on the static graph to three-dimensional space, thereby realizing the comprehensive capture of dynamic graph features in complex systems. Through multi-layer dynamic graph convolution operations, each node can perceive the information of its multi-hop spatiotemporal neighbors. Unlike the traditional method of extracting dynamic graph features by simply combining the timing module with the static graph convolution network, the lightweight dynamic graph convolution operation of this embodiment can capture richer node interaction structural features and time characteristics. This not only significantly reduces information loss, but also significantly reduces computational complexity through lightweight design, thereby improving computational efficiency and model training speed and optimizing memory usage efficiency.

[0132] Example 2

[0133] like Figure 3 As shown, a dynamic graph data feature extraction device based on a lightweight dynamic graph neural network, the device includes a dynamic graph data feature extraction method based on a lightweight dynamic graph neural network; the device also includes:

[0134] A data receiving module, used for acquiring dynamic data and converting the dynamic data into dynamic graph structure data;

[0135] It should be noted that in this embodiment, the data receiving module is a wireless or wired communication module, which is used to receive dynamic data collected by the data acquisition module, and the data acquisition module is a camera or other data acquisition device. Specifically, the data is initialized into a dynamic graph structure, that is, the data is initialized into an adjacency tensor A, whose shape is N×N×T (slice A in the tensor t represents the adjacency matrix of the graph at time t, A t The element a in the matrix ijt Indicates the connection between node i and node j in the graph at time t. If node i and node j are connected in the graph, a ijt Equal to 1; otherwise, if node i and node j are not connected in the graph, a ijt equal to 0).

[0136] An objective function module, connected to the data receiving module, for constructing a loss function using the dynamic graph structure data;

[0137] The objective function module selects the most suitable distance metric function according to the specific data pattern of the dynamic graph data to evaluate the node features in the dynamic graph and the calculation and optimization process of the feature extraction module.

[0138] A feature extraction module, connected to the data receiving module and the objective function module, for performing a convolution operation on the dynamic graph structure data to extract dynamic graph data features;

[0139] The feature extraction module combines with the objective function module to perform the dynamic graph data feature extraction process, and stores the extracted features into the feature storage unit in the storage module; it includes: an initialization unit, which initializes the relevant extraction parameters involved in the dynamic graph data feature extraction process; a lightweight dynamic graph neural network unit, which extracts dynamic graph data features according to the dynamic graph data stored in the dynamic graph data storage unit in the storage module and the extraction parameters initialized by the initialization unit; an output unit, which stores the most accurate dynamic graph data features in the lightweight dynamic graph neural network unit into the feature storage unit in the storage module.

[0140] The storage module is connected to the data receiving module and the data receiving module, and is used to store the dynamic data, dynamic graph structure data and dynamic graph data features.

[0141] The storage module stores the dynamic image data received and initialized by the data receiving module and the dynamic image data features extracted by the device; it includes a dynamic image data storage unit, which stores the dynamic image data received and initialized by the data receiving module; and a feature storage unit, which stores the dynamic image data features extracted by the device for backup.

[0142] like Figure 3 As shown, in summary, this embodiment first receives the dynamic data collected by the acquisition module through the data receiving module, and then converts these data into dynamic graph structure data. On this basis, the objective function module uses the dynamic graph structure data to construct a loss function to evaluate the calculation and optimization process of the node features and feature extraction module. The feature extraction module processes the dynamic graph structure data through convolution operations to extract key dynamic graph data features.

[0143] To further optimize performance, the feature extraction module uses a lightweight dynamic graph neural network unit that can efficiently process dynamic graph data while reducing the consumption of computing resources. The initialization unit is responsible for setting the extraction parameters, while the output unit is responsible for storing the most accurate part of the extracted features into the feature storage unit.

[0144] The storage module is an important part of the whole system. It not only stores dynamic data and dynamic graph structure data, but also is responsible for saving the extracted dynamic graph data features. The dynamic graph data storage unit and the feature storage unit jointly ensure the integrity and fast access of the data. In practical applications, this device can be applied to a variety of scenarios, such as social network analysis, traffic flow prediction, etc. By extracting dynamic graph data features, we can more accurately understand the dynamic changes in complex systems and provide strong support for decision-making.

[0145] As another optional embodiment of the present invention, optionally, the feature extraction module includes:

[0146] A parameter initialization unit, used to initialize the relevant parameters for extracting dynamic graph data features;

[0147] like Figure 3 As shown, the initialized parameters include: time conversion matrix M, the perceptual sliding window size B of the time conversion matrix, the number of lightweight dynamic graph convolution layers L, and the initialized dynamic graph feature tensor H 0 , learning rate η, regularization factor λ, current iteration number n, maximum iteration number K, termination convergence threshold ε, etc.; where:

[0148] The time transition matrix M is initialized to a lower triangular band matrix with T rows and T columns;

[0149] The initial values ​​of the perceptual sliding window size B of the time transition matrix and the number of lightweight dynamic graph convolution layers L are positive integers;

[0150] Dynamic graph feature tensor H 0 Initialize a tensor with N rows (number of nodes), F columns (feature dimension), and T time slices;

[0151] The learning rate η is initialized to a very small number, which is the initial value of the gradient descent when using the Adam optimizer during training;

[0152] The regularization factor λ is a hyperparameter that controls the degree of regularization. It can prevent the model from overfitting and is generally set to a very small number.

[0153] The current training round number n is initialized to an integer of 0, indicating that training has not yet started. During training, the value increases with the number of training rounds, indicating that the current model is in the nth training iteration process;

[0154] The maximum number of iterations K is the maximum number of training rounds, and the parameter is initialized to a large positive integer, such as 500;

[0155] The termination convergence threshold ε represents the lower limit of the error between two rounds of training. If the actual value is lower than this lower limit, training is stopped.

[0156] A lightweight dynamic graph neural network unit, connected to the parameter initialization unit, for extracting the dynamic graph data features based on the dynamic graph structure data and the initialized related parameters;

[0157] It should be noted that the prior art is that the dynamic graph feature extraction device based on the general graph convolution model generally separates the extraction process into spatial domain graph convolution and time domain information aggregation, that is, firstly, the spatial domain topological features in the graph are obtained by using the graph convolution network on a single time slice of the dynamic graph, and then the features on the continuous time slices are aggregated in the time dimension. Although the above method can effectively extract the spatiotemporal features of the dynamic graph, it inevitably destroys its integrity in the spatiotemporal domain. In addition, the existing methods still have problems such as high computational complexity and long model training time when processing large-scale data. This device implements a lightweight dynamic graph convolution module based on the tensor multiplication framework, which can fully capture the spatiotemporal features of the dynamic graph in the spatiotemporal domain. By removing the feature transformation and nonlinear activation operations in the dynamic graph convolution operation, the device significantly reduces the computational complexity of the model, thereby improving the training speed of the model and optimizing the memory usage efficiency.

[0158] The output unit is connected to the lightweight dynamic graph neural network unit and is used to extract the most accurate dynamic graph data features.

[0159] As another optional embodiment of the present invention, optionally, the storage module includes:

[0160] A dynamic graph data storage unit, used for storing the dynamic graph structure data;

[0161] The feature storage unit is used to store the dynamic graph data features.

[0162] Example 3

[0163] A computer device comprising:

[0164] processor;

[0165] a memory for storing processor-executable instructions;

[0166] Among them, the processor is configured to implement a dynamic graph data feature extraction method based on a lightweight dynamic graph neural network in Example 1 when executing the executable instructions.

[0167] 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.

[0168] Among them, the processor is used to control the overall operation of the computer device to complete the above-mentioned dynamic graph data feature extraction method based on the lightweight dynamic graph neural network.

[0169] 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.

[0170] 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.

[0171] 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.

[0172] 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.

[0173] 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 dynamic graph data feature extraction method based on lightweight dynamic graph neural network.

[0174] Although the 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 present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A dynamic graph data feature extraction method based on a lightweight dynamic graph neural network, characterized in that: The method comprises: S1, collecting dynamic data, and converting the dynamic data into dynamic graph structure data; S2. Constructing a loss function based on the dynamic graph structure data; S3, performing a convolution operation on the dynamic graph structure data to extract node features; S4, removing the feature transformation operation and the nonlinear activation operation in the convolution operation to extract dynamic graph data features; S5. Determine whether the loss value of the loss function converges or whether the number of iterations reaches an upper limit; If so, output the dynamic graph data features; If not, the loss function is iteratively optimized using stochastic gradient descent.

2. A method for extracting dynamic graph data features based on a lightweight dynamic graph neural network as claimed in claim 1, characterized in that: The expression of the loss function is: Among them, ω represents the loss value, i represents the i-th time slice graph in the dynamic graph structure data, and Y i represents the true weight label of the i-th time slice graph in the dynamic graph structure data, log() represents the logarithmic function, represents the predicted weight label of the i-th time slice graph in the dynamic graph structure data, λ represents the regularization factor, ‖ ‖ F represents the F norm, H 0 Represents the initialized dynamic graph feature tensor.

3. A method for extracting dynamic graph data features based on a lightweight dynamic graph neural network as claimed in claim 1, characterized in that: The expression of the convolution operation in step S3 is: Among them, H l+1 represents the output feature of the l+1th layer convolution operation, σ() represents the nonlinear activation function operation, Represents the symmetric normalized adjacency tensor used to perform the l-th layer of convolution, * represents the tensor multiplication operator, H l represents the output feature of the l-th layer convolution operation, W l represents the feature transformation tensor of the lth layer, ×3 represents M transformation operations, M represents the time transformation matrix, and g represents the area operation.

4. A method for extracting dynamic graph data features based on a lightweight dynamic graph neural network as claimed in claim 3, characterized in that: The method further includes symmetric normalization of the adjacency tensor of the dynamic graph structure data; the expression of the symmetric normalization is: in, represents the adjacency tensor after symmetric normalization, D represents the degree tensor constructed based on the adjacency tensor, g represents the area operation, and A represents the adjacency tensor of the dynamic graph structure data.

5. A method for extracting dynamic graph data features based on a lightweight dynamic graph neural network as described in claim 1 or 3, characterized in that: In step S4, the expression for removing the feature transformation operation and the nonlinear activation operation in the convolution operation is: Among them, H l+1 represents the output features of the l+1th layer convolution operation, represents the symmetric normalized adjacency tensor used to perform the l-th layer of lightweight dynamic graph convolution, * represents the tensor multiplication operator, H l represents the output features of the l-th layer convolution operation, ×3 represents M conversion operations, M represents the time conversion matrix, and g represents the area operation.

6. A method for extracting dynamic graph data features based on a lightweight dynamic graph neural network as claimed in claim 5, characterized in that: The expression of the M conversion operation is: in, represents the output of the l+1th layer M conversion operation, t represents the tth slice of the element of the required M conversion operation in the time dimension, k represents the index of the M conversion operation from the first slice to the tth slice in the time dimension, and m tk represents the value of k at position t in the time conversion matrix, Represents the input of the l-th layer M transformation operation.

7. A dynamic graph data feature extraction device based on a lightweight dynamic graph neural network, characterized in that: The device comprises the dynamic graph data feature extraction method based on a lightweight dynamic graph neural network according to any one of claims 1 to 6; the device also comprises: A data receiving module, used for acquiring dynamic data and converting the dynamic data into dynamic graph structure data; An objective function module, connected to the data receiving module, for constructing a loss function using the dynamic graph structure data; A feature extraction module, connected to the data receiving module and the objective function module, for performing a convolution operation on the dynamic graph structure data to extract dynamic graph data features; The storage module is connected to the data receiving module and the data receiving module, and is used to store the dynamic data, dynamic graph structure data and dynamic graph data features.

8. A dynamic graph data feature extraction device based on a lightweight dynamic graph neural network as claimed in claim 7, characterized in that: The feature extraction module comprises: A parameter initialization unit, used to initialize the relevant parameters for extracting dynamic graph data features; A lightweight dynamic graph neural network unit, connected to the parameter initialization unit, for extracting the dynamic graph data features based on the dynamic graph structure data and the initialized related parameters; The output unit is connected to the lightweight dynamic graph neural network unit and is used to extract the most accurate dynamic graph data features.

9. The dynamic graph data feature extraction device based on a lightweight dynamic graph neural network as claimed in claim 7, characterized in that: The storage module comprises: A dynamic graph data storage unit, used for storing the dynamic graph structure data; The feature storage unit is used to store the dynamic graph data features.

10. A computer device, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to implement a dynamic graph data feature extraction method based on a lightweight dynamic graph neural network as described in any one of claims 1 to 6 when executing the executable instructions.

Images