A review of rumor detection techniques in social networks

Social media rumors are rich in variety and have many more differences than traditional rumors. Social network rumors contain a lot of multimedia content such as audio and video. Content features are extracted from images and texts.

-Textual features

Scholars use natural language processing technology to mine and process the text of rumors. Since rumor texts contain words, sentences, and topics, rumor detection using the textual features of rumors can be further refined to utilize its lexical features, syntactic features and topic features. Manan Vohra and Misha- Kakkar in 2018 used the LDA algorithm for topic feature extraction and textual feature extraction for rumor detection [47]. Based on the work of different researchers, this paper collates specific models for the application of these three features.

With word-based extraction, one model uses the total number of words and characters, the total number of unique words, and the average length of words to accomplish the task of rumor detection. Other researchers have taken the sentence dimension and selected the number of keywords in the sentence, sentiment score and lexical annotation [20]. The Bag of Words (BOW) model is the best known of these [31]. Benjamin et al. used two simple BOW representations for text input: term frequency and term frequency inverse document frequency [36]. Topic features are used for rumor detection, focusing more on understanding textual information and its relationships in the corpus. For example, the potential Dirichlet distribution classifier is trained for 18 topics [26].

Rumor detection methods based on textual features are common approaches in early rumor detection, such as the pioneering researcher in rumor detection, Castillo [7]. Refine textual features into string length, number of words, whether it contains punctuation (question marks, exclamation marks, etc.), whether it includes emoticons, personal pronouns (first person, second person, third person), number of capital letters, time of publication, number of positive words, number of negative words, whether it contains tags, whether it contains links.

-Visual features

Pictures or videos may accompany rumors in social networks, and just a video or a picture contains a rumor. Therefore, the visual feature of rumors cannot be ignored entirely in the process of classifying whether the information in social networks is a rumor or not. The visual features of rumors can be divided into visual statistical features and visual content.

A rumor detection study mines time-series features from images to point out that rumor content contains outdated images [44]. Using visual content as a feature, mainly from the perspective of graphics, to explore the information contained in the picture, like the clarity and consistency of an image.

2.3.2 Social context features

-Propagation features

Takahashi et al. [20] studied the spread of Twitter. Rumors have been found to have a characteristic of exploding retweets in a short time, with much higher retweet rates than regular tweets. Most false rumors can be identified not only by the content of the false rumor but also by how people react to it and who those people are. The pattern of dissemination, combined with the subject matter of the thread and the sentiment of the user response, can be a strong indication of whether the original message is true or false. Internet rumors can be traced back to the source of propagation in the network because of their specificity. Some research has used the network features of comments and retweets for rumor detection or proliferation forms.

The pattern of information dissemination in social networks can be modelled as a tree that reflects not only the relationship between retweets and their authors but also the temporal behavior and sentiment of retweets. The most superficial characteristics of propagation in Twitter and Weibo are the number of message retweets and comments. They can be seen as a reaction behavior of users to the message. In addition to the number of retweets and comments, propagation features also include the content and timing features of comments attached to retweets. A recently emerged feature of the propagation process is sentiment consistency, which refers to whether the sentiment tendency of a post is consistent with the sentiment tendency expressed in the comments.

-User features

The most crucial factor that cannot be missing in spreading rumors on social networks is people. User features are mainly about the essential features of users in social networks [4], such as their age, gender, education. Users’ behaviors on social platforms are the behavioral features of users. These behaviors include users’ posting, liking, commenting and other behaviors, reflecting users’ activities and historical behaviors. The longer and more frequently users register on social platforms, the easier it is for them to identify the authenticity of the information. It can be seen that the more messages users post, the less likely they are to post rumors. Users’ network features refer to their relationship networks, which can be used as a basis for determining whether a user is likely to post rumors. A detailed description of the features used for social network rumor detection is shown in Table 2.

3.1.1 Available public datasets

In the decade of booming rumor detection technology, a large and rich dataset has been generated for researchers to use. This review will present some public datasets from social networks of excellent quality. A description of the publicly available rumor datasets is shown in Table 3.

-CREDBANK

A corpus designed to bridge this gap by systematically combining machine and human computation. CREDBANK contains over 60 million tweets in total [32].

-MULTI dataset

This dataset was published in 2017 and contains message posts collected from Sina Weibo between May 2012 and January 2016. It is a multimedia dataset containing images [13].

-Media Eval

This dataset contains about 9,000 rumor and 6,000 non-rumor tweets from 17 rumor-related events. This dataset can find the text content, additional images or videos, and several social contexts for each tweet [13].

-FakeNews Net

The FakeNews Net [39] is one of the most comprehensive fake news detection benchmarks.It contains 24,000 articles about content in social contexts, including user reactions and comments. It collects fake and factual news articles from fact-checking sites PolitiFact and GossipCop. The real tags of the news articles are provided by experts, ensuring the quality of tags.

-Emergent

Emergent is a new dataset from the Digital News Project for rumor debunking. It contains 300 rumors and 2,595 related news articles, collected and tagged by journalists. Each relevant article is summarized into a headline and labeled to indicate whether its position is supportive, opposing, or neutral. Emergent provides a real-world data source for various natural language processing tasks in a fact-checking context.

In the rumor detection technology research, reasonable and adequate use of existing public data sets will reduce a certain amount of workload. However, if oriented to the latest social events or social news, the existing data sets cannot meet the detection of new fields and new rumors due to the different environment and background. So at the same time, we need to combine other data collection methods to improve and enrich the data used in the training of rumor detection models. From the experience of previous studies, it can be summarized that typical other data collection methods are data collection through Application Program Interfaces (APIs) provided by social platforms and crawlers.

-APIs

In the past part of the research on rumor detection technology, researchers have used Representational State Transfer (REST) APIs provided by social platforms to collect the required web data on social networks. Currently, the main social platforms using APIs for data collection are Twitter and Sina Weibo. However, researchers are more interested in collecting textual data related to rumors on these social platforms. In the face of today’s booming short-video social platforms, no systematic rumor dataset has been generated, but they provide APIs that can collect data.

-Build crawlers to collect data

While it is convenient to use APIs for data collection, sometimes researchers facing special needs and problems also need to build their crawlers to collect data. Using a web packet capture tool, the login protocol, data request process, and the correspondence between each request uniform resource locator (URL). Data are analyzed according to the features of social platforms. Then, the program is implemented to simulate the login process to the web page by obtaining cookies and creating sessions. After the simulated login, the tweet information of the logged-in user is used as the seed set to complete data collection and analysis using the get method via Hyper Text Transfer Protocol (HTTP).

The advantages and disadvantages of different data collection methods are compared as shown in Table 4. Compared with building crawlers to collect data and collecting data through APIs provided by social platforms, the outstanding advantage is that researchers can get the data that best meets their research work. Because sometimes the APIs provided by social platforms have limitations. For example, Sina Weibo’s API can capture 2000 user comments under a tweet at once. And it cannot get the user comments collapsed among the comments, which often results in huge data loss. The crawler method simulates the process of ordinary users browsing information through the program using a client such as a browser. Using this method, researchers can obtain data that meet specific needs, and can define the format and content of the data themselves. This method is not restricted to the official collection method, and can acquire data efficiently and comprehensively. However, building a crawler also has some defects that cannot be ignored. First, the crawler technology is relatively difficult. Second, the use of crawler means to collect data may face legal risks. But the advantage of using the API provided by social platforms to collect data is that it is simple and fast.

The researchers determined the features of rumors by studying the time, structure and language of rumor propagation, and identified key differences between rumor and non-rumor propagation. It was found that random forest results would be better compared to other algorithms, with an accuracy rate of about 90% and a recall rate of about 89% [27]. Other researchers explored the task of automatically classifying rumor positions expressed on the social networks. Using a dataset of more than 4,300 manually coded tweets, they constructed a supervised random forest model for the task, achieving an accuracy of more than 88% across a variety of rumor types [58]. The researchers propose a new set of features specific to the automatic recognition problem for Twitter’s shorter texts. These features significantly improve the accuracy of the classifier and achieve results described above on a recent benchmark dataset. It enables the classification of short texts using a random forest model without known feature extraction [2].

4.2.2 Hidden Markov model

In 2015, Vosoughi S [49] tested various models to detect and verify Twitter rumors. Detecting different rumors is achieved by clustering the asserted arguments about the facts. The researchers trained two HMMs, one with observed data about false rumors and the other with observed data about true rumors.

To predict the truthfulness of rumors, Vosoughi S et al. determined the salient features of rumors by studying three aspects of information dissemination: the linguistic style of the rumor, the features of the people involved in spreading the information, and the characteristics of network transmission [48]. The predictive accuracy of the time series of these features extracted from the Twitter dataset was generated using a Hidden Markov Model (HMM). The model was trained and tested on 209 rumors (113 true, 96 rumors) and the algorithm was able to accurately predict the truthfulness of 75% of the rumors [31]. And they compared the performance of the HMM model relative to four baselines. These baselines are: a majority classifier, a forwarding classifier, an N-gram classifier, and a baseline based on The classifier trained with the features Castillo et al. [17] used, called CAST.

4.2.3 Logistic regression

The researchers used a logistic regression classifier and developed features to check titles and their consistency with statements, achieving an accuracy rate of 73% [16]. Other researchers then developed a classification model using multiclass logistic regression and showed how features other than the usual bag of words used could be helpful. Compared to an RTE system trained on the same data, the accuracy was improved by 26% [16].

4.2.4 SVM

SVM algorithms are widely used for classification and regression tasks. The support vector machine algorithm uses a straight line or a hyperplane to classify statistical data into different classes. Most studies chose Twitter as the premise of the study before 2012. Yang et al. [26] studied the trustworthiness of microblog messages for the first time. The features they chose to extract from microblogs include: content features, user features, location features and two new propagation features. The researchers first considered the first three features, trained the SVM classifier using the RBF kernel function, and then used only the account-based features. They introduced account- and location-based features into each of the three basic features to investigate their effectiveness. Using the same SVM classifier and the same RBF kernel function, the classification accuracy was improved to different degrees. The clear advantages of combining these two newly proposed features into the classification task are shown.

After this, a team of researchers trained the SVM on various content-based features and obtained an F1 score of 87 [37]. Their contribution demonstrates that discourse structure analysis is an essential method for automatic rumor detection and an effective complement to lexical-semantic analysis. In 2015, researchers constructed Radial Basis Function (RBF) kernel vectors by selecting a dataset on Sina Weibo and extracting 23 features from it after building a message propagation tree, with an accuracy of 91% for SVM [26]. But the limitation is that some of these features are specific to the microblogging platform. In 2016 another group of researchers used content-based features for SVM detection with an accuracy of 93% [8].

4.2.5 Decision tree

In 2011, Castillo selected four types of features and used the J48 decision tree classifier to implement the reliable detection of messages on Twitter with an accuracy of 89% [7]. Bodnar et al. used user features to check the trustworthiness of users through a decision tree with an an accuracy of 75% [5]. Other research groups have used six classification methods for rumor detection and found a 96% accuracy rate for rumor accuracy classification using decision trees [17].

4.2.6 LDA

The researchers put forward the features of topic type, user type, average emotion and forwarding time, and extracted the topic of the message through Dirichlet distribution. The probability distribution of the topic in the message can be obtained through Equation 1. β_1,k represents all topics from 1 to K, β _i represents the distribution of the I th theme word, θ _d represents the proportion of topics in the d th message, z_d,n represents the subject of the nth word in the d th message, and w_d,n represents the nth word in the d th message [47].

p ( β 1 , k , θ 1 , D , z 1 , D , w 1 , D ) = ∏ i = 1 k p ( β i ) ∏ d = 1 D p ( θ d ) ( ∏ n = 1 N p ( z d , n | θ d ) p ( w d , n | β 1 , k , z d , n ) )(1)

Rumor detection methods based on DL are taking shape as researchers begin to consider applying the emerging deep learning techniques to the field of social network rumor detection. More and more scholars are using recurrent neural network (RNN) [24], gate recurrent unit (GRU) [35, 51], long short term memory (LSTM) [13], convolutional neural network (CNN) [12] to conduct rumor detection. Ma et al. [24] proposed a recurrent neural network-based model (RvNN) that uses RNN to capture changes in the contextual information of comments over time. CNN first appeared in the field of computer vision, and later it can be applied to text processing after deformation. The model architecture of early rumor detection using CNN is shown in Fig. 7. Two tree structures, top-down and bottom-up, are built, with each response posting as a tree node. The response relationship between the content and semantics of the postings can be captured by recursive feature learning along the tree structure, enabling fast and accurate rumor detection. The method shows excellent rumor detection capability in the early stage. Wang et al. [52] built a sentiment lexicon and used a two-layer GRU to obtain fine-grained sentiment expressions of tweets. The rumorEval and PHEME datasets were identified by applying LSTM, which detects rumors by collecting datasets related to review users, with an F1 score of 96% [13]. Song et al. [42] treat all forwarded messages as a sequence and achieves credible early detection study by CNN. Wu et al. [26] constructed a hybrid model to detect false information. Their proposed method uses the RCNN model, which is a CNN-based model that uses a recursive structure as the convolutional layer of the CNN model.

OPEN IN VIEWER

Fig. 7

Model Architecture of Early CNN Application in Rumor Detection [12].

Traditional deep learning-based rumor detection methods eliminate the manual construction of feature engineering. However, the natural end-to-end structure makes it difficult to grasp the critical components of rumor information, and the model training lacks controllability, long training time and model complexity. Therefore, introducing attention mechanisms for rumor detection.

Chen et al. [53] proposed a deep RNN-based attention model, called CallAtRumor, which can study the temporal representatipons of consecutive posts and mine the hidden features. CallAtRumor attempts to incorporate attention mechanisms to extract implicit and explicit rumor features from repetitive, ever-changing tweets to detect selected attention-getting messages in social network message sequences. Compared with other models, CallAtRumor can handle redundant data and play a role in improving rumor detection. It achieves 88.63% and 87.10% accuracy on Twitter and Sina Weibo, respectively.

Jin et al. [23] observed that rumors usually consist of images and texts. Therefore, based on Chen’s, attention-Recurrent Neural Network (att-RNN) is proposed to extract the common features of text and social features using LSTM. Then image features are extracted using VGG-19 network. Finally, the image features are fused into the joint features of text and social features for the final rumor classification.

RNN, CNN and autoencoder-based methods have achieved significant improvements over traditional methods. However, most early deep learning-based rumor detection methods process each message independently, ignoring the correlation between notes. Social networks are a common type of graphical data representing social relationships between individuals or organizations. On social networks, there are rich structural associations between different messages. For example, once a user posts a posting, all his followers will receive the posting, while the followers may also re-post or comment on the posting. In addition to the textual information in rumors, recently proposed detection methods have started exploiting the graph structure in propagation networks for rumor detection. Graph neural networks can maintain the form of rumor propagation compared to CNN and RNN. The flow of the rumor detection method based on the propagation structure is shown in Fig. 8.

OPEN IN VIEWER

Fig. 8

Flow of rumor detection method based on propagation structure [12].

The spreading features of rumors cannot be ignored. In recent work, in order to extract the structural features of propagation, researchers proposed Graph Convolutional Networks [3, 57]. GCN and GAT have been successfully employed on rumor detection [46]. Since the RvNN model only considers the pattern of propagation and ignores the structural problem of extensive dispersion in rumor detection measurements, yet propagation and dispersion are two important structural features of rumors, and with this purpose, Bian et al. [3] apply bidirectional GCN to learning communication structure features by examining the propagation structure and diffusion structure of rumors. Each layer of GCN involves source-posted information, thus enhancing the root impact of rumors.

Liu et al. [55] construct text graph tensor to describe semantic, syntactic and sequential contextual information, and implement intra- and inter-graph propagation of information through GCN. Yao et al. [56] build graphs based on word co-occurrence and document-word relationships. Then, in order to learn words and documents, they select GCN. Zhang et al. [59] used the model proposed by Li et al. [28] to effectively achieve inductive learning of new words and capture of contextual word relations.

These methods are limited to representing rumor propagation as a static graph and are not ideal for capturing the dynamic information of rumors. Choi et al. [12] proposed a graph convolutional network with an attention mechanism, which became dynamic GCN for rumor detection in 2021. The researchers first represent rumor posts and their response posts as dynamic graphs, and temporal information is used to generate a series of graph snapshots. The attention mechanism is used to learn the representation of graph snapshots and obtain structural and temporal information of rumor propagation. The main components of the Dynamic GCN (DYNGCN) model are snapshot generation, GCN, readout layer, and an attention mechanism. They are responsible for the functions of rumor propagation representation, representation learning on graph snapshots, node embedding aggregation of global graph representations, and sequential learning from a series of graph snapshots, respectively. Upon generating the graph snapshots S = {S⁽¹⁾, S⁽²⁾, …, S^(T)} and their adjacency matrices A = {A⁽¹⁾, A⁽²⁾, …, A^(T)}, researchers conduct representation learning on the graph snapshots with the graph convolutional networks (GCNs). A^(t) ∈ R^N(t)×N(t), where N^(t) is the number of nodes in the snapshot, and feature matrix X ∈ R^N(t)×F, the learnable parameters W _k  ∈ R^{d_k-1×d _k} , where k ^th layer produce node embeddings H _k  ∈ R^{N(t)×d _k} . The GCN model used is shown in Equations (2)–(4). H k = σ ( A ˆ H k - 1 W k )(2) A ˆ = D ˆ - 1 2 A ˜ D ˆ - 1 2(3) A ˜ = A + I N , D ˜ ii = ∑ j A ˜ jj(4)

Since social networks often have various images and video messages attached to them in addition to text messages. Image information is more attractive than textual information, and they have rich visual information that can assist in rumor detection. Therefore, extracting multimodal features for multimodal information containing text and image information is also the focus of today’s research. The flow of the rumor detection method based on multi-modal features is shown in Fig. 9.

OPEN IN VIEWER

Fig. 9

Flow of rumor detection method based on multimodal features.

-Attention-Recurrent neural network

Jin et al. [23] found that with the continuous improvement of social platforms, publishers can not only upload text information, but also upload one or more pictures. Therefore, based on Chen’s work, an attention-Recurrent Neural Network (att-RNN) model is proposed to extract the joint features of text and social features using Long-Short-Term Memory. Then image features are extracted using VGG-19 network, and finally visual features are fused into the joint features of text and propagation features for the final rumor classification.

Researchers defined a tweet instance I = {T, S, V} as a tuple representing three different modalities of contents: the textual content T, the social context S, and the visual content V. The proposed model took features from these modalities (R _T , R _S , R _V ), and aimed to learn a reliable representation R _I as the aggregation of T, S and V for the given tweet I. Their proposed visual-neuron attention mechanism weights the contributions of different neurons for different words. To achieve this goal, they take the output hidden state h _m of the LSTM at each time step as guidance. h _m is connected into a fully-connected layer with non-linearity ReLU function and a fully-connected layer with softmax function to obtain the attention vector A _m  ∈ R⁵¹², which has the same dimension as the visual neurons R _V . A m = W α f 2 ψ ( W α f 1 h m )(5)

-Event adversarial neural networks

Since most existing methods have difficulty in identifying rumors of emerging different types of events, they can only mine event-specific features and cannot apply the extracted features to new types of events. To address this shortcoming, researchers proposed an EANN [50] framework for new rumor detection based on multimodal features. The researchers built a new, publicly available dataset for false news detection that is an order of magnitude larger than the previous largest publicly available faulty news dataset of its kind. CNN is used to extract texture features, while the pre-trained VGG19 [41] is used to extract visual features efficiently. After that, the features are stitched together, and finally the stitched features are input to the rumor detection module and the event classification module, respectively. The model learns the invariant event features and can detect the emerging events.

The lack of training data will lead to the inability to truly describe the real social network. Therefore, scholars have proposed transfer learning [38]. Incomplete training data is a problem that can arise during rumor detection, which can lead to the inability to identify features in features deeply and can use migration learning to improve the accuracy of the model by transferring knowledge in the relevant domain.

Yang and Zhang [54] proposed a migration learning algorithm for short text mining called Automatic Transfer Learning (AutoTL). Johnson and Zhang [25] implemented a semi-supervised framework, which can improve the accuracy of text classification. Do and Gaspers [14] greatly improved the accuracy by using additional initialization parameters for the unlabeled dataset. Recently, Guo et al. [18] proposed a deep migration model called TL-CNN, which leverages knowledge of reviews and ratings in the e-commerce domain to achieve accurate rumor detection. The architecture of the TL-CNN model is shown in Fig. 10.

OPEN IN VIEWER

Fig. 10

TL-CNN architecture.

Health information, such as food and disease, is closely related to people’s lives, so rumors in the health field account for a large part of all rumors. Especially this year, infectious diseases are spreading all over the world, and social media are full of health rumors. The researchers [40] used the health information data collected from Twitter in their research on the detection of health rumors. For the detection of Chinese health rumors, researchers used 88646 pieces of Chinese health data shared by the true and false information platforms of WeChat and Tencent, including 8972 pieces of rumors. Through the dynamic LSTM model, a stable accuracy of 0.9 can finally be achieved on the evaluation set. It is not difficult to see that for rumor detection technologies in different fields, early data preparation is crucial. In this process, we need the help of experts with knowledge in a certain field.