A cross-modal meta-learning method for early rumor detection based on small sample social big data.

Abstract

This invention relates to the field of social media content detection, and more specifically, to an early rumor detection method based on cross-modal meta-learning of social big data with small sample sizes. The method includes: acquiring multimodal information to be detected published on social media; extracting features using a multimodal feature extraction network; constructing a multimodal hidden information extraction network to extract hidden information, thereby obtaining multimodal hidden information; constructing a multimodal fusion processing network; using the multimodal fusion processing network to detect rumors; and optimizing the multimodal hidden information extraction network and the multimodal fusion processing network using a meta-learning algorithm to obtain the final detection result. This invention achieves rumor detection by extracting multimodal hidden information and performing deep fusion processing. Furthermore, this invention achieves early rumor detection with a small sample size by applying a meta-learning algorithm to the multimodal hidden information extraction network and the multimodal fusion processing network.

Description

Technical Field

[0001] This invention relates to the field of social media content detection, and more specifically, to an early rumor detection method based on cross-modal meta-learning of small-sample social big data. Background Technology

[0002] With the rapid development of personal computers and mobile phones, sharing news and obtaining information through online media platforms has become increasingly convenient, and news and information with subjective opinions are frequently shared on these platforms. However, people rarely verify the accuracy of news or information before sharing it, leading to the widespread dissemination of unverified content, including rumors. Compared to real news, rumors are often novel, making them more attention-grabbing and evoking emotional responses, thus causing them to spread faster and wider. Therefore, accurately identifying rumors among the massive amounts of shared information is a crucial research area. Mainstream social media platforms include Weibo, Xiaohongshu, and Douyin. These social media networks can send not only text content that people wish to express but also multimodal information such as images. This multimodal information not only enriches the diversity of people's expression of information but also increases the difficulty of rumor detection. Furthermore, rumors often have a small sample size and a relatively simple propagation structure in their early stages, leading to low accuracy when using general rumor detection methods for early detection.

[0003] Existing technology discloses a Weibo rumor detection algorithm and a rumor detection feature set based on gradient boosting trees. The feature set for rumor detection contains 23 features. A gradient boosting tree-based rumor detection algorithm is provided. This algorithm first constructs training samples according to the features in the feature set, and these training samples are used to train the Weibo rumor detection model. Then, multiple training iterations are performed on the training sample set to obtain multiple regression tree models. Each regression tree provides a prediction value, and the prediction values ​​from multiple regression trees are combined to obtain the final Weibo rumor detection model. During rumor detection, features of the Weibo post to be predicted are extracted according to the feature set, and the detection model is used to calculate the predicted value for the Weibo post. Based on the predicted value, the Weibo post to be predicted is determined to be either a rumor or not. However, this method still suffers from poor detection performance in the early stages of rumor dissemination when the sample size is small. Summary of the Invention

[0004] The purpose of this invention is to disclose an early rumor detection method based on small sample social big data cross-modal meta-learning, which is applicable to small samples and has better detection results.

[0005] To achieve the above objectives, this invention provides an early rumor detection method based on cross-modal meta-learning of small-sample social big data, comprising:

[0006] S1: Obtain the multimodal information to be detected published on social media, use a multimodal feature extraction network to extract features from the multimodal information to be detected, and obtain multimodal feature information;

[0007] S2: Construct a multimodal hidden information extraction network; use the multimodal hidden information extraction network to extract hidden information from multimodal feature information to obtain multimodal hidden information;

[0008] S3: Construct a multimodal fusion processing network, use the multimodal fusion processing network to detect rumors by analyzing multimodal hidden information, and obtain preliminary detection results;

[0009] S4: The multimodal hidden information extraction network and the multimodal fusion processing network are optimized using a meta-learning algorithm to obtain optimized multimodal hidden information extraction network and optimized multimodal fusion processing network. Rumor detection is performed based on the optimized multimodal hidden information extraction network and optimized multimodal fusion processing network to obtain the final detection result.

[0010] Further, step S1 includes: preprocessing the multimodal information to be detected by deleting redundant content, including at least one of the following: special characters, URL hyperlinks, HTML characters, and emoticons.

[0011] In step S1, the multimodal information to be detected includes text modality and image modality; wherein, the text modality uses different types of languages;

[0012] Let set P = {T, V1, V2, ..., V} N ,} represents the multimodal information to be detected, {T} represents the text modality, and {V1,V2,...,V N} represents the image modality; where the text modality consists of a sequence of words, represented as T={W1,W2,...,W N ,};

[0013] The multimodal feature information includes text feature information and image feature information;

[0014] The multimodal hidden information extraction network includes a text hidden information extraction network and an image hidden information extraction network.

[0015] Multimodal feature extraction networks include text feature extraction networks and image feature extraction networks;

[0016] Furthermore, a multimodal feature extraction network is used to extract features from the multimodal information to be detected, obtaining multimodal feature information, including:

[0017] Multimodal feature extraction networks include text feature extraction networks and image feature extraction networks;

[0018] VGG16 was used as the image feature extraction network to extract features from the image information, thus obtaining image feature information.

[0019] Using BERT as a text feature extraction network, we can extract features from text information and obtain text feature information:

[0020] P T =BERT{W1,W2,...,W N}

[0021] P V =VGG{V1,V2,...,V N}

[0022] Among them, P T This is the output feature information after BERT encoding, with a dimension of 393216, representing text feature information; P V This is the final layer output of the VGG model, with a dimension of 4096, representing image feature information. It combines text and image feature information as multimodal feature information.

[0023] Further, in step S2, constructing the multimodal hidden information extraction network includes:

[0024] Two Bi-LSTM bidirectional long short-term memory networks and a series of hidden layer structures are used to construct a text hidden information extraction network, which extracts hidden information from text features, including:

[0025] Reduce the dimensionality of the input text:

[0026] x = ReLU{W1P} T +b1}

[0027] Where x is the output after dimensionality transformation, W1 is the weight matrix of the activation function network, and b1 is the bias vector;

[0028] The reduced text dimensionality features are input into two Bi-LSTM bidirectional long short-term memory networks:

[0029]

[0030] output = cat(output1, output2)

[0031] Where output, output, and output2 are the outputs of the two-layer LSTM and the cascaded outputs along the feature dimension, respectively. These are the initial hidden states and cell states of the LSTM. These are the final hidden states and cell states of the LSTM;

[0032] Using two hidden layers with identical structures, taking the first layer as an example:

[0033] output layer1 1 =LayerNorm1(output)

[0034] output layer1 2 =ReLU(fc1(output) layer1 1 ))

[0035] output layer1 3 =Dropout(output) layer1 2 )

[0036] The final output of the second hidden layer is output. layer2 3 ;

[0037] Output the hidden text information through a fully connected layer:

[0038] output text =fc3(output) layer2 3 )

[0039] By processing the last fully connected layer, the final output dimension is 128.

[0040] Furthermore, constructing a multimodal hidden information extraction network includes:

[0041] An image hidden information extraction network is constructed using fully connected layers to extract hidden information from image features, including:

[0042] output3 = Linear(W2P) V +b2)

[0043] output image =Linear(W3output3+b3)

[0044] Where W2 is the weight matrix of the first fully connected layer, b2 is the bias vector, and P... V The input image features are represented by b3, which has a dimension of 4096; W3 is the weight matrix of the second fully connected layer, b3 is the bias vector, and the final output dimension is 128.

[0045] Furthermore, constructing a multimodal fusion processing network includes: using a multilayer sensing network to construct a multimodal fusion processing network.

[0046] Furthermore, a multimodal fusion processing network is used to detect rumors by analyzing multimodal hidden information, and the detection results include:

[0047] splicing multimodal hidden information:

[0048] output = cat(output) text ,output image )

[0049] Where, output text To hide information in the text, output image To hide information in an image, output is... text and output image The cascaded outputs represent multimodal hidden information;

[0050] Using two hidden layers with identical structures, taking the first layer as an example:

[0051] output layer1 1 =Linear(output)

[0052] output layer1 2 =ReLU(W4output) layer1 1 +b4)

[0053] output layer1 3 =Dropout(output) layer1 2 (p1 = 0.5)

[0054] Where W4 and b4 are the weight matrix and bias vector, respectively, and p1 is the pooling probability, representing the random separation of 50% of neurons;

[0055] The probability of a rumor is obtained through a fully connected layer:

[0056] output rumor =W6output layer2 2 +b6

[0057] Where W6 and b6 are the weight matrix and bias vector of the output layer, respectively, output layer2 2 This is the final output of the second hidden layer. rumorIt is a two-dimensional array that stores the probability of rumors and non-rumors, and the detection result is obtained by the probability of rumors.

[0058] Further, step S4 includes:

[0059] Initialize the parameters of the multimodal hidden information extraction network and the multimodal fusion processing network:

[0060] θ=(θ text ,θ image ,θ classifier )

[0061] Where, θ text ,θ image ,θ classifier These represent the initial parameters of the text hiding information extraction network, the image hiding information extraction network, and the multimodal fusion output, respectively.

[0062] Calculate the loss using the supporting dataset:

[0063]

[0064] Where CE represents the cross-entropy loss. Represents the supporting dataset, It is the average loss of the model on the support set;

[0065] Perform fast gradient updates based on the support set data:

[0066]

[0067] in, This represents the parameters that are updated rapidly and used as input for subsequent multi-step updates; α is the learning rate. It is the gradient output of the multimodal hidden information extraction network and the multimodal fusion processing network;

[0068] Calculate the loss function and update the parameters at step k:

[0069]

[0070] Where i represents the number of tasks. It is the gradient output of the i-th task step k-1 in the text.

[0071] It is the first step of the multi-step update input parameters

[0072] After completing the multi-step update of the support set in the previous step, the parameters of the multimodal hidden information extraction network and the multimodal fusion processing network need to be validated and updated meta-parameters. In this step, the loss function is calculated using the query dataset, backpropagation is performed, and finally, the meta-parameters are updated. The specific mathematical expressions are as follows:

[0073] Calculate the loss function for the query set:

[0074]

[0075] Where CE represents the cross-entropy loss. Represents the query set dataset. It is the model's average loss on the query set;

[0076] Calculate the final loss function for multiple tasks:

[0077]

[0078] Where loss q It is the average loss function value across multiple task query sets, where N is the total number of tasks;

[0079] Update meta parameters:

[0080]

[0081] Where, θ new This represents the updated meta-parameters after one step of gradient descent. It is the gradient of the loss on the query dataset with respect to the parameter θ.

[0082] Furthermore, this invention provides an early rumor detection system based on small-sample social big data cross-modal meta-learning, including:

[0083] Acquisition and Feature Extraction Module: Acquires the multimodal information to be detected published on social media, and uses a multimodal feature extraction network to extract features from the multimodal information to be detected, thereby obtaining multimodal feature information;

[0084] Hidden information extraction module: Construct a multimodal hidden information extraction network; Use the multimodal hidden information extraction network to extract hidden information from multimodal feature information to obtain multimodal hidden information;

[0085] Fusion Processing Module: Constructs a multimodal fusion processing network, uses the multimodal fusion processing network to detect rumors by analyzing multimodal hidden information, and obtains preliminary detection results.

[0086] The final detection module optimizes the multimodal hidden information extraction network and the multimodal fusion processing network using a meta-learning algorithm, resulting in optimized multimodal hidden information extraction and fusion processing networks. Rumor detection is then performed based on these optimized networks to obtain the final detection result.

[0087] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0088] This invention achieves the detection of rumors and non-rumors by extracting multimodal hidden information and performing deep fusion processing. Furthermore, by applying meta-learning algorithms to the multimodal hidden information extraction network and the multimodal fusion processing network, this invention enables early detection of rumors with a small sample size. The meta-learning algorithm focuses on effective learning using a small number of labeled or unlabeled samples; by learning from multiple tasks, it constructs learning strategies that allow for rapid adjustment of model parameters or learning of new features when facing new tasks, thereby improving learning efficiency and ultimately increasing accuracy. Attached Figure Description

[0089] Figure 1 This is a flowchart of the early rumor detection method based on small sample social big data cross-modal meta-learning, as described in Example 1.

[0090] Figure 2 This is a diagram of the multimodal feature extraction network structure described in Example 2;

[0091] Figure 3 This is a diagram of the multimodal hidden information extraction network structure described in Example 2;

[0092] Figure 4 This is a diagram of the multimodal fusion processing network structure described in Example 2;

[0093] Figure 5 This is a block diagram of the early rumor detection system based on small sample social big data cross-modal meta-learning, as described in Example 2. Detailed Implementation

[0094] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0095] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0096] Example 1:

[0097] This embodiment provides, as follows: Figure 1 The method shown is an early rumor detection method based on cross-modal meta-learning of social big data with small sample sizes, including:

[0098] S1: Obtain the multimodal information to be detected published on social media, use a multimodal feature extraction network to extract features from the multimodal information to be detected, and obtain multimodal feature information;

[0099] S2: Construct a multimodal hidden information extraction network; use the multimodal hidden information extraction network to extract hidden information from multimodal feature information to obtain multimodal hidden information;

[0100] S3: Construct a multimodal fusion processing network, use the multimodal fusion processing network to detect rumors by analyzing multimodal hidden information, and obtain preliminary detection results;

[0101] S4: The multimodal hidden information extraction network and the multimodal fusion processing network are optimized using a meta-learning algorithm to obtain optimized multimodal hidden information extraction network and optimized multimodal fusion processing network. Rumor detection is performed based on the optimized multimodal hidden information extraction network and optimized multimodal fusion processing network to obtain the final detection result.

[0102] This embodiment achieves the detection of rumors and non-rumors by extracting multimodal hidden information and performing deep fusion processing. Furthermore, this embodiment utilizes meta-learning algorithms in both the multimodal hidden information extraction network and the multimodal fusion processing network to achieve early detection of rumors with a small sample size. The meta-learning algorithm focuses on effective learning using a small number of labeled or unlabeled samples; by learning from multiple tasks, it constructs learning strategies that allow for rapid adjustment of model parameters or learning of new features when facing new tasks, thereby improving learning efficiency and ultimately increasing accuracy.

[0103] Example 2:

[0104] This embodiment further discloses information based on Embodiment 1:

[0105] Further, step S1 includes: preprocessing the multimodal information to be detected by deleting redundant content, including at least one of the following: special characters, URL hyperlinks, HTML characters, and emoticons.

[0106] In step S1, the multimodal information to be detected includes text modality and image modality; wherein, the text modality uses different types of languages;

[0107] Let set P = {T, V1, V2, ..., V} N ,} represents the multimodal information to be detected, {T} represents the text modality, and {V1,V2,...,V N} represents the image modality; where the text modality consists of a sequence of words, represented as T={W1,W2,...,W N ,};

[0108] The multimodal feature information includes text feature information and image feature information;

[0109] The multimodal hidden information extraction network includes a text hidden information extraction network and an image hidden information extraction network.

[0110] Multimodal feature extraction networks include text feature extraction networks and image feature extraction networks;

[0111] Furthermore, a multimodal feature extraction network is used to extract features from the multimodal information to be detected, obtaining multimodal feature information, including:

[0112] like Figure 2 The multimodal feature extraction network shown includes a text feature extraction network and an image feature extraction network:

[0113] VGG16 was used as the image feature extraction network to extract features from the image information, thus obtaining image feature information.

[0114] Using BERT as a text feature extraction network, we can extract features from text information and obtain text feature information:

[0115] P T =BERT{W1,W2,...,W N}

[0116] P V =VGG{V1,V2,...,V N}

[0117] Among them, P T This is the output feature information after BERT encoding, with a dimension of 393216, representing text feature information; P V This is the final layer output of the VGG model, with a dimension of 4096, representing image feature information. It combines text and image feature information as multimodal feature information.

[0118] Further, in step S2, as Figure 3 The multimodal hidden information extraction network shown includes:

[0119] Two Bi-LSTM bidirectional long short-term memory networks and a series of hidden layer structures are used to construct a text hidden information extraction network, which extracts hidden information from text features, including:

[0120] Reduce the dimensionality of the input text:

[0121] x = ReLU{W1P} T +b1}

[0122] Where x is the output after dimensionality transformation, W1 is the weight matrix of the activation function network, and b1 is the bias vector;

[0123] The reduced text dimensionality features are input into two Bi-LSTM bidirectional long short-term memory networks:

[0124]

[0125] output = cat(output1, output2)

[0126] Where output, output, and output2 are the outputs of the two-layer LSTM and the cascaded outputs along the feature dimension, respectively. These are the initial hidden states and cell states of the LSTM. These are the final hidden states and cell states of the LSTM;

[0127] Using two hidden layers with identical structures, taking the first layer as an example:

[0128] output layer1 1 =LayerNorm1(output)

[0129] output layer1 2 =ReLU(fc1(output) layer1 1 ))

[0130] output layer1 3 =Dropout(output) layer1 2 )

[0131] The final output of the second hidden layer is output. layer2 3 ;

[0132] Output the hidden text information through a fully connected layer:

[0133] output text =fc3(output) layer2 3 )

[0134] By processing the last fully connected layer, the final output dimension is 128.

[0135] Furthermore, constructing a multimodal hidden information extraction network includes:

[0136] An image hidden information extraction network is constructed using fully connected layers to extract hidden information from image features, including:

[0137] output3 = Linear(W2P) V +b2)

[0138] output image =Linear(W3output3+b3)

[0139] Where W2 is the weight matrix of the first fully connected layer, b2 is the bias vector, and P... V The input image features are represented by b3, which has a dimension of 4096; W3 is the weight matrix of the second fully connected layer, b3 is the bias vector, and the final output dimension is 128.

[0140] Furthermore, constructing a multimodal fusion processing network includes: such as Figure 4 The diagram shows a multi-modal fusion processing network constructed using a multi-layer sensing network.

[0141] Furthermore, a multimodal fusion processing network is used to detect rumors by analyzing multimodal hidden information, and the detection results include:

[0142] splicing multimodal hidden information:

[0143] output = cat(output) text ,output image )

[0144] Where, output text To hide information in the text, output image To hide information in an image, output is... text and output image The cascaded outputs represent multimodal hidden information;

[0145] Using two hidden layers with identical structures, taking the first layer as an example:

[0146] output layer1 1 =Linear(output)

[0147] output layer1 2 =ReLU(W4output) layer1 1 +b4)

[0148] output layer1 3 =Dropout(output) layer12 (p1 = 0.5)

[0149] Where W4 and b4 are the weight matrix and bias vector, respectively, and p1 is the pooling probability, representing the random separation of 50% of neurons;

[0150] The probability of a rumor is obtained through a fully connected layer:

[0151] output rumor =W6output layer2 2 +b6

[0152] Where W6 and b6 are the weight matrix and bias vector of the output layer, respectively, output layer2 2 This is the final output of the second hidden layer. rumor It is a two-dimensional array that stores the probability of rumors and non-rumors, and the detection result is obtained by the probability of rumors.

[0153] Further, step S4 includes:

[0154] Initialize the parameters of the multimodal hidden information extraction network and the multimodal fusion processing network:

[0155] θ=(θ text ,θ image ,θ classifier )

[0156] Where, θ text ,θ image ,θ classifier These represent the initial parameters of the text hiding information extraction network, the image hiding information extraction network, and the multimodal fusion output, respectively.

[0157] Calculate the loss using the supporting dataset:

[0158]

[0159] Where CE represents the cross-entropy loss. Represents the supporting dataset, It is the average loss of the model on the support set;

[0160] Perform fast gradient updates based on the support set data:

[0161]

[0162] in, This represents the parameters that are updated rapidly and used as input for subsequent multi-step updates; α is the learning rate. It is the gradient output of the multimodal hidden information extraction network and the multimodal fusion processing network;

[0163] Calculate the loss function and update the parameters at step k:

[0164]

[0165] Where i represents the number of tasks. It is the gradient output of the i-th task step k-1 in the text.

[0166] It is the first step of the multi-step update input parameters

[0167] After completing the multi-step update of the support set in the previous step, the parameters of the multimodal hidden information extraction network and the multimodal fusion processing network need to be validated and updated meta-parameters. In this step, the loss function is calculated using the query dataset, backpropagation is performed, and finally, the meta-parameters are updated. The specific mathematical expressions are as follows:

[0168] Calculate the loss function for the query set:

[0169]

[0170] Where CE represents the cross-entropy loss. Represents the query set dataset. It is the model's average loss on the query set;

[0171] Calculate the final loss function for multiple tasks:

[0172]

[0173] Where loss q It is the average loss function value across multiple task query sets, where N is the total number of tasks;

[0174] Update meta parameters:

[0175]

[0176] Where, θ new This represents the updated meta-parameters after one step of gradient descent. It is the gradient of the loss on the query dataset with respect to the parameter θ.

[0177] This embodiment achieves the detection of rumors and non-rumors by extracting multimodal hidden information and performing deep fusion processing. Furthermore, this embodiment utilizes meta-learning algorithms in both the multimodal hidden information extraction network and the multimodal fusion processing network to achieve early detection of rumors with a small sample size. The meta-learning algorithm focuses on effective learning using a small number of labeled or unlabeled samples; by learning from multiple tasks, it constructs learning strategies that allow for rapid adjustment of model parameters or learning of new features when facing new tasks, thereby improving learning efficiency and ultimately increasing accuracy.

[0178] Example 3:

[0179] This embodiment provides, as follows: Figure 2 The system shown is an early rumor detection system based on cross-modal meta-learning of social big data with small sample sizes, including:

[0180] Acquisition and Feature Extraction Module: Acquires the multimodal information to be detected published on social media, and uses a multimodal feature extraction network to extract features from the multimodal information to be detected, thereby obtaining multimodal feature information;

[0181] Hidden information extraction module: Construct a multimodal hidden information extraction network; Use the multimodal hidden information extraction network to extract hidden information from multimodal feature information to obtain multimodal hidden information;

[0182] Fusion Processing Module: Constructs a multimodal fusion processing network, uses the multimodal fusion processing network to detect rumors by analyzing multimodal hidden information, and obtains preliminary detection results.

[0183] The final detection module optimizes the multimodal hidden information extraction network and the multimodal fusion processing network using a meta-learning algorithm, resulting in optimized multimodal hidden information extraction and fusion processing networks. Rumor detection is then performed based on these optimized networks to obtain the final detection result.

[0184] This embodiment achieves the detection of rumors and non-rumors by extracting multimodal hidden information and performing deep fusion processing. Furthermore, this embodiment utilizes meta-learning algorithms in both the multimodal hidden information extraction network and the multimodal fusion processing network to achieve early detection of rumors with a small sample size. The meta-learning algorithm focuses on effective learning using a small number of labeled or unlabeled samples; by learning from multiple tasks, it constructs learning strategies that allow for rapid adjustment of model parameters or learning of new features when facing new tasks, thereby improving learning efficiency and ultimately increasing accuracy.

[0185] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A small sample based social big data cross-modal meta-learning early rumor detection method, characterized in that, include: S1: Obtain the multimodal information to be detected published on social media, use a multimodal feature extraction network to extract features from the multimodal information to be detected, and obtain multimodal feature information; S2: Construct a multimodal hidden information extraction network; use the multimodal hidden information extraction network to extract hidden information from multimodal feature information to obtain multimodal hidden information; S3: Construct a multimodal fusion processing network, use the multimodal fusion processing network to detect rumors by analyzing multimodal hidden information, and obtain preliminary detection results; S4: The multimodal hidden information extraction network and the multimodal fusion processing network are optimized using a meta-learning algorithm to obtain optimized multimodal hidden information extraction networks and optimized multimodal fusion processing networks. Rumor detection is then performed based on these optimized networks to obtain the final detection results, including: Initialize the parameters of the multimodal hidden information extraction network and the multimodal fusion processing network: wherein, respectively represent a text hidden information extraction network, an image hidden information extraction network, and initial parameters of multi-modal fusion processing. Calculate the loss using the supporting dataset: where CE represents the cross-entropy loss, represents the support dataset, is the average loss of the model on the support set; Perform fast gradient updates based on the support set data: in, This indicates a parameter for rapid updates, used as input for subsequent multi-step updates. It's the learning rate. It is the gradient output of the multimodal hidden information extraction network and the multimodal fusion processing network; Calculate the loss function and update the parameters at step k: in Indicates the number of tasks. It is the gradient output of the i-th task at the (k-1)-th step in the text. It is the first step of the multi-step update input parameters ; After completing the multi-step update of the support set in the previous step, it is necessary to validate the parameters of the multimodal hidden information extraction network and the multimodal fusion processing network, and then update the meta-learning parameters. In this step, the loss function is calculated using the query dataset, backpropagation is performed, and finally the meta-learning parameters are updated. The specific mathematical expression is as follows: Calculate the loss function for the query set: Where CE represents the cross-entropy loss. Represents the query set dataset. It is the model's average loss on the query set; Calculate the final loss function for multiple tasks: in Z is the average loss function value across multiple task query sets, where Z is the total number of tasks. Update meta parameters: in, This represents the updated meta-parameters after one step of gradient descent. It is the loss on the query dataset relative to the parameters. The gradient.

2. The method for early rumor detection based on cross-modal meta-learning of social big data with small sample sizes as described in claim 1, characterized in that, Step S1 includes: preprocessing the multimodal information to be detected by deleting redundant content, including at least one of the following: special characters, URL hyperlinks, HTML characters, and emoticons.

3. The method for early rumor detection based on small-sample social big data cross-modal meta-learning according to claim 1, characterized in that, In step S1, the multimodal information to be detected includes text modality and image modality; wherein, the text modality uses different types of languages; Let set This represents the multimodal information to be detected. Represents text modality, The image modality is represented; the text modality consists of a sequence of words, represented as follows: ; The multimodal feature information includes text feature information and image feature information; The multimodal hidden information extraction network includes a text hidden information extraction network and an image hidden information extraction network; Multimodal feature extraction networks include text feature extraction networks and image feature extraction networks.

4. The early rumor detection method based on small-sample social big data cross-modal meta-learning according to claim 3, characterized in that, In step S1, a multimodal feature extraction network is used to extract features from the multimodal information to be detected, obtaining multimodal feature information, including: Multimodal feature extraction networks include text feature extraction networks and image feature extraction networks; VGG16 was used as the image feature extraction network to extract features from the image information, thus obtaining image feature information. Using BERT as a text feature extraction network, we can extract features from text information and obtain text feature information: in, It is the output feature information after BERT encoding, with a dimension of 393216, which is text feature information; This is the final layer output of VGG16, with a dimension of 4096, containing image feature information. It combines text and image feature information as multimodal feature information.

5. The early rumor detection method based on small-sample social big data cross-modal meta-learning according to claim 3, characterized in that, In step S2, constructing the multimodal hidden information extraction network includes: Two Bi-LSTM bidirectional long short-term memory networks and a series of hidden layer structures are used to construct a text hidden information extraction network, which extracts hidden information from text features, including: Reduce the dimensionality of the input text: in, This is the output result after dimensionality transformation. It is the weight matrix of the activation function network. It is the bias vector; The reduced text dimensionality features are input into two Bi-LSTM bidirectional long short-term memory networks: in, These are the outputs of two Bi-LSTM bidirectional long short-term memory networks. For cascaded outputs along the feature dimension, These are the initial hidden state and cell state of Bi-LSTM. These are the final hidden state and cell state of Bi-LSTM; Using two hidden layers with identical structures, taking the first layer as an example: The final output of the second hidden layer is: ; Output the hidden text information through a fully connected layer: After processing the last fully connected layer, the final output dimension is 128.

6. The method for early rumor detection based on small-sample social big data cross-modal meta-learning according to claim 5, characterized in that, In step S2, constructing the multimodal hidden information extraction network includes: An image hidden information extraction network is constructed using fully connected layers to extract hidden information from image features, including: in, It is the weight matrix of the first fully connected layer. It is a bias vector. It represents image feature information with a dimension of 4096; It is the weight matrix of the second fully connected layer. It is a bias vector, and the final output size is 128 dimensions.

7. The method for early rumor detection based on cross-modal meta-learning of social big data with small sample sizes as described in claim 1, characterized in that, In step S3, constructing the multimodal fusion processing network includes: constructing the multimodal fusion processing network using a multilayer sensing network.

8. The method for early rumor detection based on small-sample social big data cross-modal meta-learning according to claim 6, characterized in that, Using a multimodal fusion processing network, rumor detection is performed on multimodal hidden information, and preliminary detection results are obtained, including: Concatenating multimodal hidden information: in, Hiding information in text Hiding information in an image yes as well as The cascaded outputs represent multimodal hidden information; Using two hidden layers with identical structures, taking the first layer as an example: in, and These are the weight matrix and the bias vector, respectively. It is the pooling probability, representing that 50% of neurons are randomly separated; The probability of a rumor is obtained through a fully connected layer: in, and These are the weight matrix and bias vector of the output layer, respectively. This is the final output of the second hidden layer. It is a two-dimensional array that stores the probability of rumors and non-rumors, and the detection result is obtained by the probability of rumors.

9. An early rumor detection system based on small-sample social big data cross-modal meta-learning, characterized in that, include: Acquisition and Feature Extraction Module: Acquires the multimodal information to be detected published on social media, and uses a multimodal feature extraction network to extract features from the multimodal information to be detected, thereby obtaining multimodal feature information; Hidden information extraction module: Construct a multimodal hidden information extraction network; Use the multimodal hidden information extraction network to extract hidden information from multimodal feature information to obtain multimodal hidden information; Fusion Processing Module: Constructs a multimodal fusion processing network, uses the multimodal fusion processing network to detect rumors by analyzing multimodal hidden information, and obtains preliminary detection results; The final detection module optimizes the multimodal hidden information extraction network and the multimodal fusion processing network using a meta-learning algorithm, resulting in optimized multimodal hidden information extraction and fusion processing networks. Rumor detection is then performed based on these optimized networks to obtain the final detection result; this includes: Initialize the parameters of the multimodal hidden information extraction network and the multimodal fusion processing network: in, These represent the initial parameters for the text hiding information extraction network, the image hiding information extraction network, and the multimodal fusion processing, respectively. Calculate the loss using the supporting dataset: Where CE represents the cross-entropy loss. Represents the supporting dataset, It is the average loss of the model on the support set; Perform fast gradient updates based on the support set data: in, This indicates a parameter for rapid updates, used as input for subsequent multi-step updates. It's the learning rate. It is the gradient output of the multimodal hidden information extraction network and the multimodal fusion processing network; Calculate the loss function and update the parameters at step k: in Indicates the number of tasks. It is the gradient output of the i-th task at the (k-1)-th step in the text. It is the first step of the multi-step update input parameters ; After completing the multi-step update of the support set in the previous step, it is necessary to validate the parameters of the multimodal hidden information extraction network and the multimodal fusion processing network, and then update the meta-learning parameters. In this step, the loss function is calculated using the query dataset, backpropagation is performed, and finally the meta-learning parameters are updated. The specific mathematical expression is as follows: Calculate the loss function for the query set: Where CE represents the cross-entropy loss. Represents the query set dataset. It is the model's average loss on the query set; Calculate the final loss function for multiple tasks: in Z is the average loss function value across multiple task query sets, where Z is the total number of tasks. Update meta parameters: in, This represents the updated meta-parameters after one step of gradient descent. It is the loss on the query dataset relative to the parameters. The gradient.

Images