A rumor detection framework based on adaptive data augmentation and adversarial training

Abstract

The application discloses a rumor detection framework based on adaptive data enhancement and adversarial training, and aims to overcome the problems that the prior art only considers structure and ignores features when facing data noise problems, and does not consider the influence of adversarial attacks on rumor detection problems. The steps of the method are as follows: 1. processing the data set; 2. using an adaptive data enhancement method to weaken the influence of noise; 3. using a graph convolutional neural network to generate a graph representation; 4. using an adversarial generation module to generate an adversarial representation to improve the robustness of the model; and 5. constructing a joint loss function to optimize the rumor detection task.

Description

Technical Field

[0001] This invention relates to a rumor detection framework on social networks, and more specifically, to a rumor detection framework for social networks based on adaptive data augmentation and adversarial training. Background Technology

[0002] With the rapid development of social media platforms, while we enjoy their convenience, rumors also spread rapidly online. A large number of widely circulated rumors mislead public perception, seriously affecting the harmony of the internet and even social stability. Therefore, there is an urgent need to develop an effective framework to detect rumors on social media and reduce the damage caused by them.

[0003] Currently, commonly used rumor detection methods can be divided into three categories: One is artificial intelligence-based detection, where platforms hand over suspicious information from social networks to experienced editors or industry experts, who then use their domain knowledge and experience to verify the information's authenticity. Another is machine learning-based methods, which manually extract features from user profiles, text content, sentiment, and other factors, and then use feature engineering to achieve the rumor detection task. The last is deep learning-based methods, which utilize CNNs, RNNs, and GCNs to mine deep semantic information from events, including user characteristics, text content, and propagation structure.

[0004] Although some research has been conducted on rumor detection, existing research still has some shortcomings, mainly in the following aspects:

[0005] 1. Many methods do not take into account the impact of noise on the data, thus reducing the performance of rumor detection.

[0006] While some existing methods utilize data augmentation to reduce the impact of noise on the data, on the one hand, most methods only augment edges, ignoring noise in the features. On the other hand, these data augmentations randomly modify edges with uniform probability, potentially removing important edges from the propagation structure. This results in suboptimal performance for rumor detection.

[0007] 2. Existing rumor detection methods do not take into account the robustness of neural networks against adversarial attacks.

[0008] Rumors in the open spaces of social media can be deliberately forwarded or responded to by malicious users or bots, thereby altering their spread and comments, which is conducive to the propagation of rumors. Furthermore, rumor-mongers intentionally misspell words or inject adversarial perturbations using high-frequency words. These artificially designed adversarial perturbations deceive neural networks, reduce the performance of rumor detection, and hinder the real-world application of rumor detection methods. Summary of the Invention

[0009] The main technical problem addressed by this invention is the lack of attention to the impact of noise on data and the lack of robustness of rumor detection methods against adversarial attacks in existing technologies. This invention proposes a rumor detection framework based on adaptive data augmentation and adversarial training.

[0010] To achieve the above objectives, the present invention employs the following technical solution:

[0011] 1. The rumor detection framework based on adaptive data augmentation and adversarial training includes the following steps:

[0012] (1) Data processing module

[0013] 1) Divide the dataset according to event type, and each event corresponds to a rumor propagation graph composed of users.

[0014] 2) Construct a rumor propagation graph and adjacency matrix based on the user IDs in the event and the relationships between users' replies and comments.

[0015] 3) Use Bidirectional Encoder Representation from Transformers (BERT) to pre-train the user's features (including some basic personal information of the user, such as: number of followers, followed tree, nickname, etc.) to obtain the feature matrix.

[0016] 4) Construct a label matrix based on the labels of the rumor events. Depending on the dataset, there may be two or four labels.

[0017] 5) Based on the results of 1), 2), 3), and 4), the dataset can be represented as follows:

[0018] C = {c1, c2, Λc} m},

[0019] c i ={{G i ,y i} i Λ},

[0020] G i =(V i E i ),

[0021] A i ={a i,st ,a i,s+1t+1, Λ},

[0022]

[0023] yi ∈{N,F,T,U} / y i ∈{T,F},

[0024] Where C represents the event set, m represents the total number of events; G represents the rumor propagation graph; V represents the node set, where each node represents a user; E represents the edge set, indicating the forwarding and commenting relationships between nodes; A represents the adjacency matrix of the rumor propagation graph, a st y is an element in the adjacency matrix A. It takes the value 1 when an edge from node s to node t belongs to edge set E; otherwise, it takes the value 0. i The labels represent events: N represents non-rumor, F represents false rumor, T represents true rumor, and U represents unverified rumor.

[0025] (2) Use an adaptive data augmentation module to enhance both edges and features.

[0026] 1) We utilize the centrality Θ of the nodes i (·) is used to calculate the centrality of the edge. The calculation formula is as follows:

[0027]

[0028] 2) After obtaining the edge centrality, in order to reduce the influence of high-density nodes on the importance of the edges, we perform a logarithmic operation on the edge centrality. Then, we use k i,st The importance of calculating edges The calculation formula is as follows:

[0029]

[0030] Where k i,max and k i,avg Representing k respectively i,st The maximum and average values ​​of ρ. e It is a hyperparameter that controls the probability of edge deletion and deordering. It is a truncation probability used to prevent excessive edge deletion and misalignment from affecting the graph structure for rumor propagation.

[0031] By employing this adaptive data augmentation method, we calculate the importance of edges and modify unimportant edges through deletion and derangement, while retaining important edges to maximize the integrity of the rumor propagation graph and improve the performance of our rumor detection framework.

[0032] 3) Use the adaptive data augmentation module to augment the features.

[0033] 4) First, we select a masking vector Q. i ∈{0,1}F F represents the dimension of the feature vector X.

[0034] 5) Then we use the masking vector Q i To generate a new feature vector, the calculation method is as follows:

[0035]

[0036] Where [·; ·] represents the operator for connecting operations, o represents the element-wise multiplication, and x j The feature vector representing node j, j∈{1,2,Λn} i}

[0037] 6) Using the obtained new feature vectors, we calculate the importance weights of each dimension of the feature. Taking a b-dimensional feature as an example, the calculation method is as follows:

[0038]

[0039] Where, |x i,jb | represents the absolute value of the b-th dimension feature of node j. Θ i (·) represents the centrality of node j.

[0040] 7) Similar to edge feature enhancement, we take the logarithm of the node centrality. Then, we use s i,st To calculate the importance of b-dimensional features The calculation method is as follows:

[0041]

[0042] Where s i,max and s i,avg They represent s respectively i,st The maximum and average values ​​of ρ. x It is a hyperparameter that controls the amplitude enhancement of nodal features. It is a truncation probability to prevent node attributes from being excessively corrupted. After adaptive data augmentation, we obtain the augmented propagation graph, adjacency matrix, and eigenvectors.

[0043] (3) Generate graph representations using a bidirectional graph convolutional neural network, the steps are as follows:

[0044] 1) Construct the original graph and the enhanced propagation graph into a top-down propagation graph and a bottom-up propagation graph, respectively.

[0045] 2) Using a bidirectional graph convolutional neural network to learn top-down and bottom-up propagation graph representations, taking the augmented propagation graph as an example, the calculation formula is as follows:

[0046]

[0047]

[0048] in, and X′ represents the adjacency matrix of the top-down augmentation graph and the adjacency matrix of the bottom-up augmentation graph, respectively. i It is the enhanced feature matrix. and σ represents the learnable parameters of a k-layer graphical convolutional neural network. σ represents the ReLU activation function. and These represent top-down and bottom-up augmented graph representations, respectively.

[0049] 3) We perform average pooling operations on both the top-down and bottom-up augmented graph representations, and then aggregate the results to obtain the augmented graph representation H′. i The calculation formula is as follows:

[0050]

[0051]

[0052]

[0053] Here, MEAN(·) represents the average pooling operation. ⊕ represents the matrix addition operation.

[0054] 4) Perform steps 2) and 3) on the original graph to obtain the original graph representation H. i .

[0055] (4) Adversarial training of the graph representation is performed using both Fast Gradient Method (FGM) and Projection Gradient Descent (PGD). The steps are as follows:

[0056] 1) For the FGM method, we first calculate the backpropagation gradient g, as follows:

[0057]

[0058] Where λ(·) represents the loss function, φ represents the learnable weight parameters, and H′ i The graph representation after aggregation, taking the augmented graph as an example, y i Represents a label.

[0059] 2) We use the gradient g obtained in 1) to generate an adversarial perturbation and add it to the augmented graph representation. The calculation steps are as follows:

[0060]

[0061] H′ i,atk =H′ r +r atk ,

[0062] Where, r atk Representing the adversarial perturbation, γ is a hyperparameter controlling the magnitude of the perturbation, ||·|| 2 Representing the l2 paradigm. We utilize the adversarial representation H′ obtained. i,atk Let's update the weight parameter φ.

[0063] 3) For the PGD method, we divide the final adversarial perturbation generation process into multiple small steps, and set the total number of small steps as T. First, we calculate the gradient g in small step t. t+1 Generate adversarial examples Then, Add to the diagram representation H′ i In the middle, the generation representation H′ i,atk The calculation method is as follows:

[0064]

[0065]

[0066] 4) We utilize the learnable weights φ obtained through FGM or PGD attack methods. atk By using fully connected layers and normalization layers to combine the original graph representation and the enhanced graph representation, an adversarial original graph representation z is generated. i,atk And adversarial enhancement graph representation z′ i,atk The calculation method is as follows:

[0067] z i,atk =Norm((H i φ atk +b0)φ atk +b1,

[0068] z′ i,atk =Norm((H′) i φ atk +b0)φ atk +b1,

[0069] Where b represents deviation.

[0070] 5) We concatenate the original representation and the enhanced graph representation with the adversarial original graph representation and the adversarial enhanced graph representation, respectively, and then perform a final concatenation of the concatenated representations to obtain the final graph representation. The calculation method is as follows:

[0071] zi =concat(z i,atk H i ),

[0072] z′ i =concat(z′) iatk H′ i ),

[0073] z i,sum =concat(z′) i ,z i ),

[0074] Here, concat(·) represents the join function. i,sum The representative finally stated.

[0075] (5) Finally, we use the loss function to optimize our rumor detection task, and the steps are as follows:

[0076] 1) Calculate the cross-entropy loss function.

[0077] 2) Calculate the contrastive loss function.

[0078] 3) Construct a joint loss function by adding the cross-entropy loss function and the contrastive loss function, calculated as follows:

[0079] loss=(1-β)*loss entroy +β*loss conv

[0080] Where, loss entroy This represents the cross-entropy loss function, a widely used classification loss function that utilizes label information to improve classification performance. conv The contrastive loss function brings samples with the same label closer together and separates samples with different labels, thus improving the detection performance and robustness of the framework. β is a hyperparameter that controls the ratio of the two loss functions.

[0081] Compared with the prior art, the beneficial effects of the present invention are:

[0082] 1. Events are affected by noise during propagation. Existing methods only consider the utilization of propagation structure information, neglecting the impact of noise on structure and content. However, our framework performs adaptive data augmentation on both structure and features, thus preserving important structure and features and enhancing the diversity and randomness of the data. Therefore, it improves the framework's robustness to noise.

[0083] 2. Rumors in the open space of social media can be injected with adversarial interference by malicious users, altering their propagation structure and comment content. These artificial adversarial perturbations are less noticeable and more destructive than noise. While some methods consider the uncertainty of propagation, they do not account for the impact of adversarial attacks on rumor detection. Our framework utilizes the FGM / PGD method for adversarial training to simulate real adversarial attacks in a high-dimensional space, allowing the framework to learn their features during the training phase. Furthermore, the generated adversarial representations are also used as hard samples; combined with contrastive learning, our framework is forced to learn invariant basic representations from these hard samples. This improves the framework's robustness to adversarial attacks and its performance in rumor detection.

[0084] In summary, this invention provides a framework for rumor detection using adaptive data augmentation and adversarial training methods. We utilize adaptive data augmentation to improve the model's robustness to noise, while adversarial training and contrastive loss further enhance the model's stability and robustness. Attached Figure Description

[0085] Figure 1 This is a schematic diagram showing the functions and connections of the various modules of the computer program for the rumor detection framework based on adaptive data augmentation and adversarial training described in this invention.

[0086] Figure 2 This is a flowchart of the rumor detection framework based on adaptive data augmentation and adversarial training described in this invention.

[0087] Figure 3 This is a flowchart illustrating the process by which the data processing module processes the dataset, extracts key information, and obtains input data within the rumor detection framework based on adaptive data augmentation and adversarial training as described in this invention.

[0088] Figure 4 This is a flowchart illustrating the process by which the adaptive data augmentation module generates an augmented graph within the rumor detection framework based on adaptive data augmentation and adversarial training, as described in this invention.

[0089] Figure 5 This is a flowchart illustrating the process of generating graph representations of the original graph and the augmented graph in the bidirectional graph convolution module of the rumor detection framework based on adaptive data augmentation and adversarial training described in this invention.

[0090] Figure 6 This is a flowchart illustrating the process of generating adversarial representations of the original and augmented graphs in the adversarial training module of the rumor detection framework based on adaptive data augmentation and adversarial training described in this invention.

[0091] Figure 7This is a flowchart illustrating the process of generating adversarial representations of the original and augmented graphs in the adversarial training module of the rumor detection framework based on adaptive data augmentation and adversarial training described in this invention. Detailed Implementation

[0092] The present invention will now be described in detail with reference to the accompanying drawings:

[0093] The rumor detection framework based on adaptive data augmentation and adversarial training described in this invention aims to address the problem of data augmentation of features, which is currently lacking in technological attention. Existing methods cannot select appropriate edges and features for data augmentation, affecting data integrity, and also lack stability against malicious user interference. This invention proposes an adaptive data augmentation method to improve the stability of the framework and an adversarial training method to simulate malicious user attack behavior, thereby improving the model's stability.

[0094] See Figure 1 To achieve the goal of a rumor detection framework based on adaptive data augmentation and adversarial training, a computer program was developed, which includes five functional modules: a data processing module, an adaptive data augmentation module, a bidirectional graph convolutional network, an adversarial training module, and a joint loss function module. The functions of each module are as follows:

[0095] 1. Data Processing Module:

[0096] The data processing module includes constructing the dataset into an event set, where each event contains an event propagation graph and tag data. It then constructs an adjacency matrix of the propagation graph using the relationships between comments and reposts among users in the dataset, and constructs a feature matrix using basic user information (nickname, number of followers, number of fans, etc.). The adjacency matrix and feature matrix are then used as input.

[0097] (1) Construction of user adjacency matrix

[0098] By leveraging the relationships between user comments and replies, user IDs are randomized into numbers, and then edge relationships between users are generated using these user IDs. An adjacency matrix of the propagation graph is constructed based on these edge relationships.

[0099] (2) Pre-trained feature matrix

[0100] The feature matrix of the propagation graph is constructed using the user's basic information. To better construct the feature matrix, the BERT pre-training method is used to train the feature matrix.

[0101] (3) Obtain input data

[0102] The resulting adjacency matrix and characteristic matrix are used as inputs for the next module.

[0103] 2. Adaptive Data Augmentation Module

[0104] (1) Perform adaptive data augmentation on edges

[0105] We calculate the importance of edges and use this information to perform deordering or deletion operations on less important edges, while preserving the most important edges. This also ensures the integrity of the graph and enhances the robustness of the model.

[0106] (2) Adaptive data augmentation of features

[0107] We calculate the importance of each dimension's features, and then mask the unimportant dimensions to remove the unimportant features while retaining the important dimensions, thus ensuring the integrity of the features.

[0108] (3) How to determine important edges or features

[0109] To calculate the importance of edges, we compare the importance of each edge with a threshold. If the importance of an edge is lower than the threshold, we consider it to be of low importance and perform deordering or deletion operations on low-importance edges. To calculate the importance of features, we compare the importance of each feature in each dimension with a threshold. If the importance of a feature is lower than the threshold, we perform a masking operation on the feature to increase the importance of the model.

[0110] 3. Bidirectional graph convolution module

[0111] (1) Constructing original / enhanced graphs from top-down and bottom-up perspectives.

[0112] Because the direction of rumor propagation is a crucial feature in rumor detection, we consider both bottom-up and top-down propagation directions. We consider the initial adjacency matrix of the original / enhanced graph to represent the top-down propagation direction of rumors. Since each rumor detection event is posted by a user, and then commented on and forwarded by other users, the initial adjacency matrix represents the top-down propagation direction when constructing the propagation graph in this way. We then consider the transpose of the initial adjacency matrix as the bottom-up propagation direction. Therefore, we obtain the adjacency matrices of the original / enhanced graph for both propagation directions.

[0113] (2) Obtain the graph representations of the original graph and the enhanced graph.

[0114] We pass the original and augmented graphs, one from top to bottom and the other from bottom to top, through two layers of graph convolutional neural networks (Graph Convolutional Neural Networks). The graph convolution operation on the top-down original and augmented graphs is called passing through the TDGCN layer, and the graph convolution operation on the top-down original and augmented graphs is called passing through the BUGCN layer. Through these two layers, we obtain the original graph representation and the augmented graph representation, respectively, from top to bottom and bottom to bottom. We then concatenate the top-down and bottom-down original graph representations to obtain the final original graph representation, and similarly, we concatenate the top-down and bottom-down augmented graph representations to obtain the final augmented graph representation.

[0115] 4. Adversarial Training Module

[0116] (1) Update weights using FGM / PGD methods

[0117] We use two attack methods, FGM and PGD, and the final original graph representation and augmented graph representation we obtained to obtain gradients, and continuously update our weights by updating gradients.

[0118] (2) Construct fully connected layers and normalized layers using updated weights.

[0119] We utilize the weights updated through the attack method as the weights for the fully connected layer and the normalized layer. We then use the weights and the final original graph representation and the enhanced graph representation to generate adversarial representations and enhanced adversarial representations.

[0120] (3) Obtain the final representation

[0121] We concatenate the original graph representation and the adversarial graph representation to obtain the final original overall graph representation. Simultaneously, we concatenate the augmented graph representation and the augmented adversarial graph representation to obtain the final augmented overall graph representation. Finally, we concatenate the final original overall graph representation and the final augmented overall graph representation together to obtain the final representation.

[0122] 5. Joint Loss Function Module

[0123] (1) Constructing the contrastive loss function

[0124] We use the learned final representation as an anchor point. Then, we treat representations with the same label as the anchor point as positive samples and representations with different labels as negative samples. We use a cosine function to shorten the distance between positive samples and widen the distance between negative samples. The calculation method is as follows:

[0125]

[0126] Where, loss conv S represents the comparative loss. v It is a set of events, where i is an event c. iThe index, and event c i It is the anchor point, and its final representation is z. i,sum j is the index of the event, which is ultimately represented as z. j,sum It shares the same label as the anchor point and is a positive sample. k is the index of the event, ultimately represented as z. k,sum It has a different label from the anchor point and is a negative sample. p (i) represents the positive sample set, S n (i) represents the negative sample set. cos(·) represents the cosine similarity, and τ is the temperature coefficient, which controls the shape of the distribution.

[0127] (2) Constructing the cross-entropy loss function

[0128] Cross-entropy loss is a widely used classification loss function that utilizes label information to improve classification performance. We construct the cross-entropy loss function using the predicted and true labels, and the calculation formula is as follows:

[0129]

[0130] Where m represents the number of tags, CA represents the category of the tags, and y i,ca Representative event c i The true label, This represents the predicted label.

[0131] (3) Constructing the joint loss function

[0132] A joint loss function is constructed using the cross-entropy loss function and the contrastive loss function, and the calculation method is as follows:

[0133] loss=(1-β)*loss entroy +β*loss conv

[0134] Here, β is a hyperparameter, which is a tradeoff coefficient that controls the ratio of the two loss functions.

Claims

1. A rumor detection framework based on adaptive data augmentation and adversarial training, comprising the following steps: (1) Processing data according to actual problems; 1) Divide the data set according to the event type, and each event corresponds to a rumor propagation graph composed of users; 2) According to the ID of the user in the event and the relationship between the user and the reply and comment, construct the rumor propagation graph and the adjacency matrix; 3) Pre-train the user features using Bidirectional Encoder Representation from Transformers (BERT), including the number of followers, the number of followed, and the nickname, to obtain the feature matrix; 4) Construct a label matrix according to the labels of rumor events; Depending on the data set, the label may be two or four; 5) Represent the data set according to the results of 1), 2), 3), and 4): ， ， ， , wherein a set of representative events, a total number of representative events; a representative rumor propagation graph; a set of representative nodes, each node representing a user; a set of representative edges, indicating the existence of forwarding and commenting relationships between nodes; an adjacency matrix representing the rumor propagation graph, is an element in the adjacency matrix , when an edge from node to node belongs to the edge set , it takes the value of 1, otherwise it takes the value of 0; a label representing an event, representing a non-rumor, representing a false rumor, representing a true rumor, representing an unverified rumor; (2) Use the adaptive data augmentation module to enhance the edges and features; 1) Using the centrality of the nodes to calculate the centrality of the edges , the formula is as follows: 2) After obtaining edge centrality, in order to reduce the influence of high-density nodes on the importance of edges, logarithmic operation is performed on the edge centrality ; then, the importance of the edge is calculated by using The calculation formula is as follows:​ where and represent the maximum and average values of respectively; is a hyper-parameter to control the probability of edge deletion and misplacement; is a truncation probability to prevent the over-deletion and misplacement of edges from affecting the structure of rumor propagation graph. Through this adaptive data augmentation method, the importance of the edge is calculated, and the unimportant edge is modified by deleting and misarranging, and the important edge is preserved to maximize the integrity of the rumor propagation graph and improve the performance of the rumor detection framework; 3) Use the adaptive data augmentation module to enhance the data of the features 4) First, a masking vector is selected , represents the dimension of the masking vector; 5) Then the new feature vector is generated using the masking vector as follows: wherein represents an operation connector, represents a dot product between elements, represents a node characteristic vector, ; 6) Using the new feature vector obtained, the importance weight of each dimension of the feature is calculated to Taking the feature vector as an example, the calculation method is as follows: wherein, representative node first absolute value of the eigenvalue; representative node centrality; 7) Similar to edge feature augmentation, log of node centrality ; then, the importance of the features of the dimension is calculated , as follows:​​ wherein and respectively represent the maximum and average values of is a hyperparameter that controls the amplitude of the node feature enhancement; is a truncation probability that prevents over-destroying of node attributes; and the enhanced propagation graph, the adjacency matrix, and the feature vector are obtained after adaptive data enhancement. (3) Use a bidirectional graph convolutional neural network to generate a graph representation, with the following steps: 1) Construct the original graph and the enhanced propagation graph into a top-down propagation graph and a bottom-up propagation graph, respectively; 2) Use a bidirectional graph convolutional neural network to learn the top-down propagation graph representation and the bottom-up propagation graph representation, for example, the propagation graph, and the calculation formula is as follows: ， ， wherein, and represent the adjacency matrix of the top-down augmented graph and the adjacency matrix of the bottom-up augmented graph, respectively; is the augmented feature matrix, and represent the learnable parameters of the layer graph convolutional neural network; represents the activation function of the RELU; and represent the top-down augmented graph representation and the bottom-up augmented graph representation, respectively; 3) Perform average pooling operation on the top-down enhanced graph representation and the bottom-up enhanced graph representation respectively, and aggregate the results obtained together to obtain an enhanced graph representation The calculation formula is as follows: ， ， ， wherein, represents an average pooling operation; represents a matrix addition operation; 4) Perform steps 2) and 3) on the original graph to obtain the original graph representation; (4) Use Fast Gradient Method (FGM) and Projection Gradient Descent (PGD) to perform adversarial training on the graph representation, with the following steps: 1) For the FGM method, first compute the backpropagation gradient in the following way: wherein, represents a loss function, represents a learnable weight parameter, represents an aggregated graph representation, here exemplified by an augmented graph, represents a label; 2) Using the gradient obtained in 1) , generate an adversarial perturbation and add it to the augmented image representation, the computation steps are as follows: ， ， wherein, represents an adversarial perturbation, is a hyperparameter controlling the size of the perturbation, represents a norm; the adversarial representation is used to update the weight parameters ; 3) For the PGD method, the final adversarial perturbation generation process is divided into multiple small steps and the total number of small steps is set to , first, according to the gradient calculated in the small step , an adversarial sample is generated; then, the is added to the graph representation , a representation is generated, and the calculation is as follows: 4) using the learnable weights obtained by the FGM or PGD attack method , using the full connection layer and the normalization layer to generate the adversarial original graph representation and the adversarial enhanced graph representation from the original graph representation and the enhanced graph representation in the following manner: ， ， wherein representative bias; 5) Concatenate the original graph representation , the adversarial original graph representation , the enhanced graph representation , and the adversarial enhanced graph representation to obtain the final graph representation, in the following way: ， ， ， wherein represents a linking function; represents a final representation; (5) Finally, use the loss function to optimize the rumor detection task, with the following steps: 1) Construct a contrastive loss function Take the final representation as an anchor point, then take the representations with the same label as the anchor point as positive samples, and take the representations with different labels as negative samples, use the cosine function to narrow the distance between positive samples and distant negative samples, the calculation method is as follows: . wherein, represents the contrastive loss, is a set of events, is an event with index i, while event is an anchor point, which is finally represented as ; is an event with index i, which is finally represented as , which has the same label as the anchor point, is a positive sample; is an event with index i, which is finally represented as , which has a different label from the anchor point, is a negative sample; represents the positive sample set, represents the negative sample set; represents the cosine similarity, is a temperature coefficient, which controls the shape of the distribution; 2) Construct a cross-entropy loss function Cross-entropy loss is a widely used classification loss function that uses label information to improve classification performance. Use the predicted label and the true label to construct a cross-entropy loss function, the calculation formula is as follows: wherein, representing a number of labels, representing a class of labels, representing an event a true label, representing a predicted label; 3) Construct a joint loss function Use the cross-entropy loss function and the contrastive loss function to construct a joint loss function, the calculation method is as follows: wherein, is a hyper-parameter, is a trade-off coefficient that controls the proportion of the two loss functions.

Images