A method for detecting fake news in social media based on meta-transfer learning

Abstract

The present invention discloses a method for detecting fake news on social media based on meta-transfer learning, which collects text and image data of an event on social media, extracts text features and image features using a deep learning model, and uses a meta-transfer learning model to learn knowledge in historical events to help the model detect newly occurring events. The core mechanism of the present invention is a two-stage training process of meta-learning: in the event adaptive training stage, the training set is divided into different tasks to train the internal model and optimize the initialization parameters of the model. In the specific event detection stage, the optimized model is used to test the newly occurring event, and the model is quickly optimized for the specific event. The method of the present invention can extract common knowledge from different events for migration to the detection of new events, and can achieve accurate new event fake news detection using a small amount of new event sample data, so as to have the ability to quickly detect newly occurring events.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning, and in particular relates to a method for detecting fake news on social media. Background Art

[0002] "Fake news" refers to social posts or news that have a certain factual basis and can be proven to be false within a limited time. The rapid development of social media platforms has given any user the freedom to post fake news, so fake news can spread quickly and widely among the crowd, not only causing social panic, but also burying hidden dangers to people's health. It is of great significance to effectively detect fake news in the social media space.

[0003] Traditional fake news detection methods mainly combine the content features of the message (such as text and comment information), propagation features (such as forwarding structure) and interaction features (such as the number of likes), and apply deep neural networks such as Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) to extract features and classify and identify them. However, traditional methods require a large amount of labeled data to train the model. For new events, it is often difficult to collect high-quality relevant labeled data required for training the model in a short period of time, resulting in poor model generalization performance. In addition, different fake news events have different topics and user comment content, and often present different data feature distributions. The existing detection methods ignore the distribution differences of true and false news for different events. The model mostly learns high-order features for specific events, and the recognition effect for new events that have not appeared in the training set is usually poor. Existing fake news detection methods have poor generalization and cannot quickly detect new events. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention provides a social media fake news detection method based on meta-transfer learning, collects text and image data of an event on social media, uses a deep learning model to extract text features and image features, and uses a meta-transfer learning model to learn knowledge in historical events to help the model realize the detection of newly occurring events. The core mechanism of the present invention is a two-stage training process of meta-learning: in the event adaptive training stage, the training set is divided into different tasks to train the internal model and optimize the initialization parameters of the model. In the specific event detection stage, the optimized model is used to test the newly occurring events, and the model is quickly optimized for specific events. The method of the present invention can extract common knowledge from different events for migration to the detection of new events, and accurate new event fake news detection can be achieved using a small amount of new event sample data, so as to have the ability to quickly detect newly occurring events.

[0005] The technical solution adopted by the present invention to solve the technical problem includes the following steps:

[0006] Step 1: Use the Text-CNN network to extract the text features of the message, and use the VGG-19 network to extract the image features of the message; for the context data x of event e e,k ∈X e , which is input into the feature extractor as raw data, and the extracted feature embedding is represented as c e,k ∈C e ;X e Represents the context data of the event, C e represents the high-dimensional features learned after inputting the context data into the feature extractor;

[0007] Step 2: Represent the label as an embedding of the same dimension as the image features and text features, i.e., label embedding, and concatenate it with the extracted feature embedding:

[0008]

[0009] In the formula, v e,k represents the label vector corresponding to the k-th data, K represents the number of messages in the data set, V e Represents a set of label vectors;

[0010] Step 3: Represent the context embedding as The feature vector of the detected target is expressed as The two vectors are concatenated and input into the true and false message classifier, and the output vector is d represents the dimension of the input vector;

[0011] Step 4: Use the label embeddings vec(fake) and vec(real) as metrics and compare them with the output of the true and false message classifier O e,i Perform similarity measurement:

[0012]

[0013] Use softmax to map the similarity measure vector to [0,1] as the probability representation of whether the message is true or false text;

[0014] Step 5: For each event e, construct two data sets: support set and query set In the event adaptation training phase, the support set is used Training is performed with the i-th message As the detection target, the rest of the data As context data, the context data of the detected target is input into the model and the prediction result is input; the loss is calculated by comparing the predicted label and the actual label To calculate:

[0015]

[0016] Where θ represents all parameters in the model;

[0017] After calculating the loss, perform gradient descent for each event e to update the parameters θ:

[0018]

[0019] Step 6: Specific event detection phase, using the parameter set θ obtained in the event adaptive training phase e As initial parameters, the query set and the entire support set As input, we get the label set predicted by the query set. Then compare it with the actual label set of the query set to calculate the loss:

[0020]

[0021] Step 7: Through the meta-learning process, an optimal initialization parameter set is learned, which learns to detect news events using the given context input as a condition.

[0022] The beneficial effects of the present invention are as follows:

[0023] The method of the present invention can extract common knowledge from different events for migration to the detection of new events, and can realize accurate new event false news detection using a small amount of new event sample data, so as to have the ability to quickly detect newly emerging events. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The figure is a flow chart of the method of the present invention.

[0025] Figure 2 Schematic diagram of adaptive training of the method of the present invention.

[0026] Figure 3 The loss curve of the model after training on the Weibo dataset and the validation curves of different indicators of the embodiment of the present invention. (a) Training process loss, (b) validation process curve. DETAILED DESCRIPTION

[0027] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0028] The present invention proposes a classification method for meta-transfer learning. For each event e, two data sets are constructed for it: support set and query set The goal of this invention is to use the knowledge gained from past events and some example samples about event e (such as support set ), for the corresponding query set Seek an effective prediction function F(s)→{0,1} so that the prediction result satisfies:

[0029]

[0030] Specifically, the present invention includes three modules: the feature extractor is divided into two parts, one part is a text feature extractor, whose input is the unstructured message text content, and the text content feature vector is extracted by using the Text-CNN network; the other part is an image feature extractor, which uses the VGG19 network to extract the feature vector of the input image. For the support set, in order to capture the semantics of the label, the true or fake label is converted into a semantic vector representation by using the label embedding algorithm. The feature aggregator receives the label embedding and the learned high-dimensional vector to avoid the information loss caused by the dimensional difference between the two. The detector performs classification detection of true and false messages, and optimizes the entire model using the gradient descent algorithm. In the training stage, the present invention utilizes the training process of meta-learning, and the entire training stage is divided into two parts: the first part uses the support set and the target set for training, and the loss of each batch is added as the total loss for gradient descent, and finally a best initialization parameter is obtained. The second part is about the fine-tuning of the model for the target event, and the training set and the test set are conventionally used for training and parameter optimization. The present invention can be combined with user interaction data, and the meta-transfer learning method can be used to detect fake news in social media based on small samples and multimodality, and quickly detect new events.

[0031] To achieve the above tasks, the present invention adopts the following technical solutions:

[0032] like Figure 1 As shown, a social media fake news detection method based on meta-transfer learning includes the following steps:

[0033] Step 1: Use the Text-CNN network to extract the text features of the message, and use the VGG-19 network to extract the image features of the message. For the context data x of event e e,k ∈X e , which is input into the feature extractor as raw data, and the extracted features are represented as c e,k ∈C e .

[0034] Step 2: To preserve the classification information contained in the label variable, the label is represented as an embedding with the same dimension as the image feature and text feature, i.e., label embedding. And concatenated with the extracted feature embedding:

[0035]

[0036] Step 3: Represent the context embedding as The feature vector of the detected target is expressed as The two vectors are concatenated and input into the true and false message classifier, and the output vector is

[0037] Step 4: Use the label embeddings vec(fake) and vec(real) as metrics and compare them with the output of the true and false message classifier O e,i Perform similarity measurement:

[0038]

[0039] Use softmax to map the two-dimensional vector representing similarity to [0, 1] as the probability representation of whether the message is true or false text.

[0040] Step 5: Figure 2 As shown, in the event adaptation training phase, the support set Training is performed with the i-th message As the detection target, the rest of the data As context data, the context data of the detected target is input into the model and the prediction result is input. The loss can be calculated by comparing the predicted label and the actual label To calculate:

[0041]

[0042] Where θ represents all the parameters in the model. After calculating the loss, gradient descent is performed for each event e to update the parameters θ:

[0043]

[0044] Step 6: Specific event detection phase, using the parameter set θ obtained in the event adaptive training phase e As initial parameters, the query set and the entire support set As input, we get the label set predicted by the query set. Then compare it with the actual label set of the query set to calculate the loss:

[0045]

[0046] Step 7: Through the meta-learning process, a better set of initialization parameters can be learned, which can quickly learn to use the given context input as a condition to detect the news event that has just occurred. The method ends. Specific embodiment:

[0048] Step 1: Collect post data of fake news events and determine the ID of each event.

[0049] Step 2: Use Text-CNN to extract semantic features and use VGG-19 to extract image features.

[0050] Step 3: Concatenate the semantic features and image features obtained in step 2, input them into the true or false message classifier and obtain the output vector.

[0051] Step 4: Measure the similarity between the output in step 3 and the label embedding of the true / false message to predict whether the current message is a false message.

[0052] Step 5: Use the support set to repeatedly train the model to obtain an optimal set of initialization parameters.

[0053] Step 6: For the specific event that needs to be detected, the parameter set obtained in step 5 is used as the initialization parameter for training to obtain a detection model for the specific event.

[0054] Figure 3 The loss curve of the model after training on the Weibo dataset and the validation curves of different indicators. (a) Training process loss, (b) validation process curve.

[0055] The present invention is a meta-transfer learning fake news detection method for social media, which can extract common knowledge from different events for migration to the detection of new events. It can achieve accurate new event fake news detection using a small amount of new event sample data, so as to have the ability to quickly detect newly emerging events.

Claims

1. A method for detecting fake news on social media based on meta-transfer learning, characterized in that: The following steps are involved: Step 1: Use the Text-CNN network to extract the text features of the message, and use the VGG-19 network to extract the image features of the message; for the context data x of event e e,k ∈Xe, and input it into the feature extractor as raw data. The extracted feature embedding is represented as c e,k ∈Ce; Xe represents the context data of the event, and Ce represents the high-dimensional features learned after the context data is input into the feature extractor; Step 2: Represent the label as an embedding of the same dimension as the image features and text features, i.e., label embedding, and concatenate it with the extracted feature embedding: In the formula, v e,k represents the label vector corresponding to the k-th data, K represents the number of messages in the data set, V e Represents a set of label vectors; Step 3: Represent the context embedding as The feature vector of the detected target is expressed as The two vectors are concatenated and input into the true and false message classifier, and the output vector is d represents the dimension of the input vector; Step 4: Use the label embeddings vec(fake) and vec(real) as metrics and compare them with the output of the true and false message classifier O e,i Perform similarity measurement: Use softmax to map the similarity measure vector to [0,1] as the probability representation of whether the message is true or false text; Step 5: For each event e, construct two data sets: support set and query set In the event adaptation training phase, the support set is used Training is performed with the i-th message As the detection target, the rest of the data As context data, the context data of the detected target is input into the model and the prediction result is input; the loss is calculated by comparing the predicted label and the actual label To calculate: Where θ represents all parameters in the model; After calculating the loss, perform gradient descent for each event e to update the parameters θ: Step 6: In the specific event detection phase, the parameter set θe obtained in the event adaptive training phase is used as the initial parameter to transform the query set and the entire support set As input, we get the label set predicted by the query set. Then compare it with the actual label set of the query set to calculate the loss: Step 7: Through the meta-learning process, an optimal initialization parameter set is learned, which learns to detect news events using the given context input as a condition.

Images