A rumor detection method, device and medium based on graph neural architecture search

Abstract

The present invention discloses a rumor detection method, device and medium based on graph neural architecture search, belonging to the field of Internet technology. The method includes: constructing a search space; continuously relaxing the search space to construct a hypernetwork, creating a mixed candidate operation parameterized by architecture parameters for each edge on the hypernetwork; initializing the parameters of the hypernetwork; alternately reducing the verification loss and the training loss by gradient descent to optimize the model parameters and architecture parameters of the hypernetwork; when optimizing to the pruning round, selecting an edge in the hypernetwork for pruning according to the edge selection criterion, and fixing the architecture parameters of the edge, and no longer optimizing; after determining that all edges on the hypernetwork have been optimized, obtaining the network architecture; training the obtained network architecture, and obtaining a pre-trained model for rumor detection after training. The present invention improves the efficiency of detection and the applicability of the model in different scenarios by automatically designing the characteristics of information propagation through learning the optimal neural architecture.

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to a rumor detection method, device, and medium based on graph neural architecture search. Background Art

[0002] With the growing development of the internet, people are increasingly accessing information through social networks. Social platforms have significantly increased the speed and volume of information dissemination and have become an integral part of people's lives. However, while social networks offer numerous benefits, they also create conditions for the spread of rumors. Rumors often cause significant harm to those who spread them, misleading the public, exacerbating social panic and triggering a crisis of trust. Therefore, timely identifying and debunking rumors on social networks is a crucial task. However, relying solely on manual rumor detection within massive amounts of information consumes significant manpower and resources, is inefficient, and is prone to lags. Therefore, research on automatic rumor detection methods is crucial to curbing the spread of rumors.

[0003] Many studies have focused on identifying rumors based on content features, such as linguistic and emotional characteristics. However, rumors are often deliberately fabricated, with content features very similar to those of real information. Therefore, identifying rumors based solely on content is insufficient. In social networks, the spread of information triggers user feedback and interaction, forming a propagation network. As a social feature of the source information, the propagation network helps identify its authenticity. In recent years, numerous studies have utilized information propagation networks for rumor detection. These methods have demonstrated their effectiveness by extracting features through various neural network models. With the development of graph neural networks, they have demonstrated significant advantages in graph data processing and are more suitable for learning structural features within propagation networks. Consequently, an increasing number of studies are applying graph neural networks to rumor detection.

[0004] However, the graph neural network architectures applicable to different datasets often vary significantly. Current research has designed various network architectures to learn propagation characteristics. However, real-world data scenarios are complex and diverse, and a network architecture designed manually for a particular dataset may not be effectively applicable to new scenarios. Designing a neural network for each dataset requires significant human effort, is inefficient, and lacks flexibility. With the development of neural architecture search, its effectiveness has been demonstrated in many fields. Therefore, a graph neural architecture search method specifically designed for rumor detection can be developed to automatically design the optimal neural network architecture. This allows for more flexible and efficient learning of propagation network characteristics, thereby improving detection efficiency and the model's applicability across diverse scenarios. Summary of the Invention

[0005] In order to at least solve one of the technical problems existing in the prior art to a certain extent, the purpose of the present invention is to provide a rumor detection method, device and medium based on graph neural architecture search.

[0006] The first technical solution adopted by the present invention is:

[0007] A rumor detection method based on graph neural architecture search, comprising the following steps:

[0008] Constructing a search space; candidate operations in the search space include node aggregation in classic graph neural networks, source information-guided node feature enhancement, and skip connection operations;

[0009] Continuously relax the search space to construct a hypernetwork, creating hybrid candidate operations parameterized by the architecture parameters for each edge on the hypernetwork;

[0010] Randomly initialize the parameters of the hypernetwork;

[0011] Alternately reduce validation loss and training loss through gradient descent to optimize the model parameters and architecture parameters of the hypernetwork;

[0012] When the optimization reaches the pruning round, an edge is selected in the hypernetwork for pruning based on the edge selection criteria. That is, the mixed operation is replaced by the candidate operation with the highest probability on the edge, and the architectural parameters of the edge are fixed and no further optimization is performed.

[0013] After ensuring that all edges on the hypernetwork are optimized, the network architecture is obtained;

[0014] The obtained network architecture is trained, and a pre-trained model for rumor detection is obtained after training.

[0015] Furthermore, the classic graph neural network includes: Graph Convolution Network (GCN), Graph Attention Network (GAT), Graph Sampling and Aggregation Network (Graph SAGE) or Graph Isomorphism Network (GIN).

[0016] Furthermore, the source information-guided node feature enhancement method includes: a gating mechanism, a source information enhancement mechanism, and a linear transformation, and each mechanism includes two node fusion methods: splicing and summation.

[0017] Furthermore, for the node v on the information propagation graph i The expressions for feature enhancement are:

[0018] 1) Gating mechanism:

[0019]

[0020]

[0021] 1.1) Splicing and fusion:

[0022]

[0023] 1.2) Sum and Fusion:

[0024]

[0025] Where, represents the gate vector of layer l, and represents the trainable weight parameters, Represents node v i In the original feature vector of the lth layer, c represents the root node, Represents the feature vector of the root node at layer l, Represents node v i The feature vector after passing the gating mechanism at layer l, Represents node v i In the final representation vector at layer l, ⊙ represents the Hadamard product, and || represents the concatenation operation;

[0026] 2) Source information enhancement mechanism:

[0027] 2.1) Splicing and fusion:

[0028]

[0029] 2.2) Sum and Fusion:

[0030]

[0031] 3) Linear transformation:

[0032]

[0033] Where, Represents trainable weight parameters.

[0034] Furthermore, the continuously relaxing the search space to construct a supernetwork includes:

[0035] Construct a directed acyclic graph as a super network, where each node represents a hidden layer in the super network, and each hidden layer can accept the output of all previous layers as input; each edge e i,j It represents the mixing operation performed from layer i to layer j. The mixing operation is calculated as:

[0036]

[0037] Where O represents the set of all candidate operations, represents the weight parameter of operation o, and σ represents the sigmoid function.

[0038] Furthermore, the optimization of the architecture parameters of the hypernetwork is formulated as a two-level optimization problem:

[0039]

[0040]

[0041] Where, and They represent training loss and validation loss respectively. The loss function is the cross entropy function. The approximate gradient algorithm proposed by DARTS is used for alternating optimization of gradient descent. w is the network parameter inside the neural network, α is the architecture parameter of the network architecture, and min α Indicates that under α The minimum value, argmin w Indicates The value of w when it reaches the minimum.

[0042] Furthermore, the side selection criteria include:

[0043] 1) Edge determinism:

[0044] For edge e on the hypernetwork i,j , the weight parameters of the candidate operations it contains are considered as a distribution, namely:

[0045]

[0046] Where p (i,j) For edge e i,j The weight distribution of candidate operations on , O is the set of candidate operations, For edge e i,j The weight of a candidate operation on is the weight of other candidate operations in the candidate operation set;

[0047] The uncertainty of the distribution is measured by entropy, and the weight distribution p is calculated (i,j) The entropy of is:

[0048]

[0049] Based on the calculated entropy, we further define the edge e i,j The certainty is:

[0050] EC(i,j) =exp(-H(p (i,j) ))

[0051] 2) Trend stability:

[0052] For edge e on the hypernetwork i,j , in the current optimization round T, the weight parameter for each candidate operation o The historical values ​​within the window size K are fitted with a polynomial to obtain the function of round t

[0053]

[0054] In the formula, ~ represents the function fitting relationship;

[0055] The stability of the function trend is expressed by the absolute value of the second-order derivative. i,j All candidate operations in the edge e are calculated by summing up the absolute values ​​of the second-order derivatives of their weight parameters at round T to represent the edge e i,j The trend stability is calculated as follows:

[0056]

[0057] 3) Side selection criteria:

[0058] For edge e on the hypernetwork i,j , whose score is:

[0059] score i,j =EC (i,j) ×TS (i,j)

[0060] score i,j The edge with the largest value is pruned.

[0061] Furthermore, the training of the obtained network architecture includes:

[0062] Input the feature matrix of the information propagation graph into the obtained network architecture to learn and update node features;

[0063] A global feature vector is obtained through the maximum pooling operation;

[0064] Input the global feature vector into the fully connected network layer and calculate the cross entropy loss;

[0065] The cross entropy loss is minimized by gradient descent, and the training ends when the preset optimization round is reached.

[0066] The second technical solution adopted by the present invention is:

[0067] A rumor detection device based on graph neural architecture search, comprising:

[0068] at least one processor;

[0069] at least one memory for storing at least one program;

[0070] When the at least one program is executed by the at least one processor, the at least one processor implements the above method.

[0071] The third technical solution adopted by the present invention is:

[0072] A computer-readable storage medium stores a program executable by a processor, wherein the program executable by the processor is used to perform the method described above when executed by the processor.

[0073] The present invention has the following advantages and effects compared to the prior art:

[0074] 1) The rumor detection method based on graph neural architecture search, disclosed in this paper, constructs a search space specifically tailored to the rumor detection task. By adding node enhancement operations, the model focuses on highly relevant content and filters out off-topic noise. Compared with conventional graph neural architecture search spaces, this method better meets the task requirements and can improve the performance of the architecture.

[0075] 2) The rumor detection method based on graph neural architecture search disclosed in the present invention designs a search strategy with a pruning mechanism, which can speed up the search and reduce the consumption of computing resources while ensuring the search effect.

[0076] 3) The rumor detection method based on graph neural architecture search disclosed in the present invention can adaptively search for the optimal network architecture on different data sets, reduce the consumption caused by manually designing the neural network architecture, improve efficiency and be more flexible. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following introduction is made to the drawings of the embodiments of the present invention or the related technical solutions in the prior art. It should be understood that the drawings introduced below are only for the convenience of clearly describing some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative work.

[0078] Figure 1 is a schematic diagram of a super network in an embodiment of the present invention;

[0079] Figure 2 is a schematic diagram of a hypernetwork pruning process in an embodiment of the present invention;

[0080] Figure 3 This is a flowchart of a rumor detection method based on graph neural architecture search in an embodiment of the present invention. DETAILED DESCRIPTION

[0081] The 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 are not to be construed as limiting the present invention. The step numbers in the following embodiments are provided for ease of explanation only and do not limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0082] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they cannot be understood as limitations on the present invention.

[0083] In the description of the present invention, the meaning of "several" is one or more, the meaning of "many" is two or more, and "greater than", "less than", "exceed", etc. are understood to exclude the number itself, and "above", "below", "within", etc. are understood to include the number itself. If there is a description of the first and the second, it is only for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features. In addition, "and / or" describes the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship.

[0084] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, and connecting should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0085] like Figure 3 As shown, this embodiment provides a rumor detection method based on graph neural architecture search, including the following steps:

[0086] S1. Construct a search space; the candidate operations in the search space include node aggregation in classic graph neural networks, source information-guided node feature enhancement, and skip connection operations.

[0087] In some embodiments, classic graph neural networks include: Graph Convolution Network (GCN), Graph Attention Network (GAT), Graph Sampling and Aggregation Network (Graph SAGE), and Graph Isomorphism Network (GIN).

[0088] The source information-oriented node feature enhancement method includes: gating mechanism, source information enhancement mechanism, linear transformation, and two node fusion methods: splicing and summation. i The expressions for feature enhancement are:

[0089] 1) Gating mechanism:

[0090]

[0091]

[0092] 1.1) Splicing and fusion:

[0093]

[0094] 1.2) Sum and Fusion:

[0095]

[0096] in, represents the gate vector of layer l, and represents the trainable weight parameters, Represents node v i In the feature vector of the lth layer, c represents the root node, ⊙ represents the Hadamard product, Represents node v i The final representation vector at layer l, || represents the concatenation operation;

[0097] 2) Source information enhancement mechanism:

[0098] 2.1) Splicing and fusion:

[0099]

[0100] 2.2) Sum and Fusion:

[0101]

[0102] 3) Linear transformation:

[0103]

[0104] Where, Represents trainable weight parameters.

[0105] S2. The search space is constructed into a super-network through continuous relaxation, and a hybrid candidate operation parameterized by the architecture parameters is created for each edge in the super-network.

[0106] Construct a directed acyclic graph as a super network, where each node represents a hidden layer in the super network, and each hidden layer can accept the output of all previous layers as input; each edge e i,j represents the blending operation performed from layer i to layer j. The blending operation is calculated as:

[0107]

[0108] Among them, O represents the set of all candidate operations, represents the weight parameter of operation o, and σ represents the sigmoid function.

[0109] S3. Randomly initialize the parameters of the hypernetwork.

[0110] S4. Alternately reduce the validation loss and training loss through gradient descent to optimize the model parameters and architecture parameters of the hypernetwork.

[0111] The optimization of architecture parameters can be formulated as a two-level optimization problem:

[0112]

[0113]

[0114] in, and They represent training loss and validation loss respectively. The loss function is the cross entropy function, and the approximate gradient algorithm proposed by DARTS is used for alternating optimization of gradient descent.

[0115] S5. When the optimization reaches the pruning round, an edge is selected in the super network for pruning according to the edge selection criterion. That is, the mixed operation is replaced by the candidate operation with the highest probability on the edge, and the architectural parameters of the edge are fixed and no further optimization is performed.

[0116] As an optional implementation, the edge selection criteria are as follows:

[0117] 1) Edge determinism:

[0118] For edge e on the hypernetwork i,j , the weight parameters of the candidate operations it contains are considered as a distribution, namely:

[0119]

[0120] The uncertainty of the distribution is measured by entropy, and the weight distribution p is calculated (i,j) The entropy of is:

[0121] H(p (i,j) )=-∑ o∈O p (i,j) log(p (i,j) ) (12)

[0122] Based on the calculated entropy, we further define the edge e i,j The certainty is:

[0123] EC (i,j) =exp(-H(p (i,j) )) (13)

[0124] 2) Trend stability:

[0125] For edge e on the hypernetwork i,j , in the current optimization round T, the weight parameter for each candidate operation o The historical values ​​within the window size K are fitted with a polynomial to obtain the function of round t

[0126]

[0127] The stability of the function trend is expressed by the absolute value of the second-order derivative. i,j All candidate operations in the edge e are calculated by summing up the absolute values ​​of the second-order derivatives of their weight parameters at round T to represent the edge e i,j The trend stability is calculated as follows:

[0128]

[0129] 3) Side selection criteria:

[0130] For edge e on the hypernetwork i,j , whose score is:

[0131] score i,j =EC (i,j) ×TS (i,j) (16)

[0132] score i,j The edge with the largest value is pruned.

[0133] S6. Repeat steps S4-S5 until the operations of all edges on the supernetwork are determined and the network architecture is obtained.

[0134] S7. The trained network architecture.

[0135] In some embodiments, step S7 specifically includes the following steps:

[0136] S71. Input the feature matrix of the information propagation graph into the obtained network architecture to learn and update node features;

[0137] S72, obtaining a global feature vector through a maximum pooling operation;

[0138] S73, input the global feature vector into the fully connected network layer and calculate the cross entropy loss;

[0139] S74. Minimize the cross entropy loss by gradient descent, and the training ends when the specified optimization round is reached.

[0140] S8. After the training is completed, a pre-trained model for rumor detection is obtained.

[0141] The above method is explained in detail below with reference to the accompanying drawings and specific embodiments.

[0142] The super network topology of this embodiment is as follows Figure 1 As shown in (a), each node X (L) Represents the hidden layer in the network, and L represents the number of layers. Figure 1 (b) is an example of a layer L = 2. The directed edge e from hidden layer i to hidden layer j i,j Represents the mixed operation performed on the input feature vector, which is the weighted sum of all candidate operations o(x) in the search space O Right now:

[0143]

[0144] Among them, σ represents the sigmoid function, and the weight of the mixing operation is composed of a vector of dimension O Parameterization. Candidate operations include node aggregation operations from classic graph neural networks, source-information-guided node feature enhancement operations, and skip connections. Classic graph neural networks include graph convolutional neural networks, graph attention networks, graph sampling and aggregation networks, and graph isomorphism networks. Source-information-guided node feature enhancement methods include gating mechanisms, source information enhancement mechanisms, linear transformations, and two node fusion methods: concatenation and summation.

[0145] The network parameters of the constructed hypernetwork are randomly initialized. The goal of architecture search is to learn a set of architecture parameters α={α (i,j)}, at the end of the search, by using the operation with the highest probability, i.e. Replace Mix Operation This results in a discrete architecture where each layer can receive no more than two inputs. The optimization process of the architecture parameters can be formulated as a two-level optimization problem:

[0146]

[0147]

[0148] in, and Denote the training loss and validation loss respectively, and the cross entropy loss function is used. The optimization goal is to find the validation loss Minimize the architectural parameter α * , where the weight w * is obtained by minimizing the training loss, i.e. The gradient value is calculated according to the approximate gradient proposed by the classic DARTS algorithm, so that the architecture parameter α and the network parameter w are alternately optimized using the gradient descent method.

[0149] When the optimization reaches the manually specified pruning round, an edge e is selected in the hypernetwork according to the edge selection criterion. i,j To perform pruning operations, the process is as follows Figure 2 As shown. Replace the mixed operation on the selected edge with the candidate operation with the highest probability, and fix the architecture parameter α of the edge (i,j) , no further optimization is performed. Each pruning round performs a pruning operation on the super network until all edges in the entire network architecture are determined and the final architecture is obtained.

[0150] After obtaining the final network architecture, the feature matrix of the information propagation graph in the dataset is input into it. Node features are learned and updated, and a global feature vector of the propagation graph is obtained through a maximum pooling operation. This global feature vector is then input into the fully connected network layer and optimized using gradient descent. The network architecture is trained by minimizing the cross-entropy loss between the predicted and true label distributions. Training ends after a specified number of rounds, resulting in a pre-trained model.

[0151] In summary, the present invention proposes a rumor detection method based on graph neural architecture search, which automatically designs the optimal neural architecture to learn the characteristics of information propagation, while improving the efficiency of detection and the applicability of the model in different scenarios. The characteristics of the method of the present invention are: the search space not only includes the node aggregation operation in the classic graph neural network, but also adds a source information-oriented node feature enhancement method; the search strategy designs a pruning mechanism based on gradient descent to improve the search efficiency. It should be noted that the present invention is a complementary framework to the existing rumor detection model based on graph neural network, which can be combined with the existing model to provide a more flexible and efficient new optimization idea.

[0152] This embodiment also provides a rumor detection device based on graph neural architecture search, including:

[0153] at least one processor;

[0154] at least one memory for storing at least one program;

[0155] When the at least one program is executed by the at least one processor, the at least one processor implements Figure 3 The method shown.

[0156] A rumor detection device based on graph neural architecture search in this embodiment can execute a rumor detection method based on graph neural architecture search provided by an embodiment of the method of the present invention, can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0157] The present application also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs Figure 3 The method shown.

[0158] This embodiment also provides a storage medium storing instructions or programs that can execute a rumor detection method based on graph neural architecture search provided by an embodiment of the method of the present invention. When the instructions or program are run, any combination of implementation steps of the method embodiment can be executed, and the corresponding functions and beneficial effects of the method can be obtained.

[0159] In some optional embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiment presented and described in the flow chart of the present invention is provided in an exemplary manner for the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operation and logic flow presented herein. Optional embodiments are contemplated in which the order of the various operations is changed and the sub-operations described as a part of a larger operation are performed independently.

[0160] Furthermore, although the present invention is described in the context of functional modules, it should be understood that, unless otherwise indicated, one or more of the functions and / or features described may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It will also be understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present invention. More specifically, given the properties, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be understood within the ordinary skill of an engineer. Therefore, a person skilled in the art using ordinary skill will be able to implement the present invention set forth in the claims without undue experimentation. It will also be understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present invention, which is determined by the full scope of the appended claims and their equivalents.

[0161] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0162] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0163] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0164] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0165] In the above description of this specification, reference to the terms "one embodiment / example," "another embodiment / example," or "certain embodiments / examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0166] While 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.

[0167] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A rumor detection method based on graph neural architecture search, characterized in that: The following steps are involved: Constructing a search space; wherein the candidate operations in the search space include node aggregation in classic graph neural networks, source information-guided node feature enhancement, and skip connection operations; Continuously relaxing the search space to construct a hypernetwork, creating hybrid candidate operations parameterized by the architecture parameters for each edge on the hypernetwork; Randomly initialize the parameters of the hypernetwork; Alternately reduce validation loss and training loss through gradient descent to optimize the model parameters and architecture parameters of the hypernetwork; When the optimization reaches the pruning round, an edge is selected in the hypernetwork for pruning according to the edge selection criteria, and the architecture parameters of the edge are fixed and no further optimization is performed; After ensuring that all edges on the hypernetwork are optimized, the network architecture is obtained; The obtained network architecture is trained, and a pre-trained model for rumor detection is obtained after the training; The source information-guided node feature enhancement method includes: a gating mechanism, a source information enhancement mechanism, and a linear transformation, each mechanism including two node fusion methods: splicing and summing; For node v on the information diffusion graph i The expressions for feature enhancement are: 1) Gating mechanism: 1.1) Splicing and fusion: 1.2) Sum and Fusion: In the formula, represents the gate vector of layer l, and represents the trainable weight parameters, Represents node v i In the original feature vector of the lth layer, c represents the root node, represents the feature vector of the root node at layer l, Represents node v i The feature vector after passing through the gating mechanism at layer l, Represents node v i In the final representation vector at the lth layer, ⊙ represents the Hadamard product, and || represents the concatenation operation; 2) Source information enhancement mechanism: 2.1) Splicing and fusion: 2.2) Sum and Fusion: 3) Linear transformation: Where W i (l) Represents trainable weight parameters; The side selection criteria include: 1) Edge determinism: For edge e on the hypernetwork i,j , the weight parameters of the candidate operations it contains are considered as a distribution, namely: In the formula, p (i,j) For edge e i,j The weight distribution of candidate operations on , O is the set of candidate operations, For edge e i,j The weight of a candidate operation on is the weight of other candidate operations in the candidate operation set; The uncertainty of the distribution is measured by entropy, and the weight distribution p is calculated. (i,j) The entropy of is: Based on the calculated entropy, we further define edge e i,j The certainty is: EC (i,j) =exp(-H(p (i,j) )) 2) Trend stability: For edge e on the hypernetwork i,j , in the current optimization round T, the weight parameter for each candidate operation o The historical values ​​within the window size K are fitted with a polynomial to obtain the function of round t In the formula, ~ represents the function fitting relationship; The stability of the function trend is expressed by the absolute value of the second-order derivative. i,j All candidate operations in the edge e are calculated by summing up the absolute values ​​of the second-order derivatives of their weight parameters at round T to characterize the edge e i,j The trend stability is calculated as follows: 3) Side selection criteria: For edge e on the hypernetwork i,j , whose score is: score i,j =EC (i,j) ×TS (i,j) Score i,j The edge with the largest value is pruned.

2. According to claim 1, a rumor detection method based on graph neural architecture search is characterized in that: The classic graph neural network includes: graph convolutional neural network, graph attention network, graph sampling and aggregation network or graph isomorphism network.

3. According to claim 1, a rumor detection method based on graph neural architecture search is characterized in that: The step of continuously relaxing the search space to construct a hypernetwork includes: Construct a directed acyclic graph as a hypernetwork, where each node represents a hidden layer in the hypernetwork, and each hidden layer can accept the output of all previous layers as input; each edge e i,j It represents the mixing operation performed from the i-th layer to the j-th layer. The mixing operation is calculated as: In the formula, O represents the set of all candidate operations, represents the weight parameter of operation o, and σ represents the sigmoid function.

4. According to claim 1, a rumor detection method based on graph neural architecture search is characterized in that: The architectural parameter optimization of the hypernetwork is formulated as a two-level optimization problem: In the formula, and They represent training loss and validation loss respectively, w is the network parameter inside the neural network, α is the architecture parameter of the network architecture, min α Indicates that under α The minimum value, argmin w Indicates The value of w when it reaches the minimum.

5. The rumor detection method based on graph neural architecture search according to claim 1 is characterized in that: The training of the obtained network architecture comprises: Input the feature matrix of the information propagation graph into the obtained network architecture to learn and update node features; A global feature vector is obtained through the maximum pooling operation; Input the global feature vector into the fully connected network layer and calculate the cross entropy loss; The cross entropy loss is minimized by gradient descent, and the training ends when the preset optimization round is reached.

6. A rumor detection device based on graph neural architecture search, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method according to any one of claims 1 to 5.

7. A computer-readable storage medium storing a program executable by a processor, characterized in that: The processor-executable program is used to perform the method according to any one of claims 1 to 5 when executed by the processor.

Images