Information diffusion describes the way in which information is disseminated through a network (Rogers et al., 2014). Although the original meaning of the process encompasses interpersonal communication channels (Katz and Lazarsfeld, 2017), given the increasing popularity of social media in the last two decades, information diffusion, and especially rumor propagation nowadays, tends to be linked with the dissemination of information over specific digital social networking platforms, such as Twitter. The Diffusion of Innovations (DOI) theory (Rogers et al., 2014) states that information diffusion is not solely determined by attributes of the statement or the novelty introduced, but it is also integrally related to the communication channel properties, i.e. the individuals involved or other network characteristics. Considering the accelerated pace of information diffusion in the digital age, it becomes readily apparent that rumor propagation and fake news dissemination is developing into a critically complex issue, with many and convoluted parameters.

Three techniques have been mostly employed in tackling this issue: structural analysis, temporal analysis, and content-based techniques. Structural analysis refers to the analysis of information cascades and the communication channel properties, whereas temporal analysis takes into account the transitional characteristics of information diffusion and aims to arrive at conclusions by evaluating the diffusion activity based on time. Content analysis is the most commonly used technique; by capitalizing on Natural Language Processing (NLP) advances in the last decade, it manages to offer contextual characteristics to the information diffusion problem. An essential aspect of these techniques is their ability to generalize for claims, i.e. whether the extracted results are appropriate to make a generalized claim. Content analysis provides some clear comparative advantages over the other two techniques, but it is location-specific and dependent on the ad hoc regional and social context (Siwakoti et al., 2021); thus, it does not generalize easily for several social media in various countries. On the other hand, a structural or temporal analysis of information diffusion offers universal results, which are location-independent and language-independent.

Significant research effort has been spent studying and modeling information diffusion on social media (Jin et al., 2013, 2014; Cannarella and Spechler, 2014; Goel et al., 2016). Fake news propagation and classification have received significant attention in the past few years and have been analyzed with a variety of techniques that include propagation-based, content-based and social context-based characteristics. However, existing work takes into account content information, which makes the analyses dependent on the language and the social context of users. This is also due to the scarcity of relevant datasets, which leads to results based on the few available datasets, mainly Twitter (Ma et al., 2017). In contrast, our work focuses on analysis that does not consider language and social context characteristics, enabling a location and language independent fake rumor detection.

In this paper, we focus on the problem of rumor modeling and extraction. Considering that most literature adopts the term “fake news,” in our work, we adopt the term “rumor” because we differentiate rumors from fake news as follows: we consider that rumors are doubtful statements that cannot be easily verified, while fake news is intentionally fabricated information presented as true. The difference, although subtle, distinguishes the data samples on which our methodology applies, from the generic ones that appear in the literature as samples of misinformation data, in general.

In our work, we perform temporal analysis through epidemiological models over the information diffusion on Twitter and we extract specific temporal features; the choice of Twitter is due to the availability of suitable datasets. Additionally, we analyze the graph structure of the information cascade of the communication channel and we extract graph-based features. Having extracted the features using these two techniques, we proceed to evaluate them in three classification schemes: (i) using only epidemiological features, (ii) using only graph-based features, and (iii) using a combination of epidemiological and graph-based features. More specifically, the contributions of our work are the following:

We show that epidemiological models fit rumor propagation data, especially model SEIZ, and we evaluate their performance on fake rumor classification tasks on Twitter datasets.

The paper is organized as follows. Section Related Work presents prior work. In section A Location-Independent and Language-Independent Approach, we present the epidemiological models and the graph modeling method as well as the employed classification schemes; for the epidemiological models we describe how they are employed in fitting the rumor propagation data and for the graph models how feature extraction is accomplished. Section Experiments and Evaluation presents our evaluation datasets and the results of rumor classification performance using epidemiological features, graph-based features and the combination of both. Section Conclusions presents our conclusions and directions for future work.

Epidemiological models have been used extensively to model information diffusion in complex networks. Such models classify the human population into different compartments and define different transitions between them to simulate the spread of information. The SI (Susceptible-Infected) model was originally proposed in 2001 (Pastor-Satorras and Vespignani, 2021) indicating that epidemiological models can help to describe the propagation of information on scale-free networks. Later another variant, the SIS (Susceptible-Infected-Susceptible) model was introduced (Newman, 2003) and used multiple times (Gross et al., 2006; Jin et al., 2013, 2014); the model allows infected users to return to the Susceptible compartment. Other more complex models have been proposed and used, such as the SEIR (Susceptible-Exposed-Infected-Removed) model (Wang et al., 2014), the S-SEIR model (Xu et al., 2013), where the spread of information is dependent on its value to the user, and the SCIR model (Xuejun, 2015) where a Contacted compartment is added, modeling how followers of a certain user react after he posts an online message. Another approach (Cannarella and Spechler, 2014) considers a modification of the SIR model where the adoption of an idea is considered an infection and its abandonment is considered a recovery. A more widely used model is the SEIZ (Susceptible-Exposed-Infected-Skeptics) model (Bettencourt et al., 2006); this model can fit long term ideas adoption. Recently, the SEIZ was employed to model the spread of fake and real news on Twitter (Jin et al., 2013, 2014). The authors proposed a simple method to classify news as either true or fake using the results of the SEIZ fitting; while the model performs well, the authors have used only a small amount of viral news stories to test their hypothesis.

Another promising and popular research direction related to the propagation of rumors and fake news is the exploitation of the diffusion graphs. Typically, graphs are constructed, representing information diffusion paths, and then features are extracted, which are used either to interpret the propagation, or to perform classification based on a scheme. In the following sections, we discuss how graphs and graph characteristics are employed to address the information diffusion problem. There have been several approaches aiming to provide a qualitative context in the way that information is disseminated. By studying the diffusion patterns in a social network along with the significance of individual nodes (users), one may recognize specific individuals as critically important to the information diffusion (Valente, 1996). Such users, who may be characterized as “opinion leaders,” tend to be the connection between mass media and people in the community (Katz and Lazarsfeld, 2017). Furthermore, they are more important than average individuals in the diffusion of information (Watts and Dodds, 2007) and have been linked with the size and the structural virality of the diffusion (Goel et al., 2016; Meng et al., 2018). Along the same lines of research, efforts have also been directed to understand the clustering patterns of news propagation (González-Bailón and Wang, 2013). Analysis of such clusters has yielded a deeper comprehension of the importance of “brokers,” or users that connect otherwise separated clusters. Such qualitative interpretations of the information diffusion graph structure (nodes and edges) are crucial to understand how diffusion efficiency is related to the network users (Valente and Fujimoto, 2010). Other research efforts, usually directed toward detection or classification schemes, focus more on extracting structural graph features relevant to diffusion. While there are attempts that concentrate on graph characteristics, such as the size of the cascade (graph), the root outgoing degrees, the followers count, or the geo-coordinates (Taxidou and Fischer, 2014), most directions employ a combination of temporal, graph, and/or content analysis (Abulaish et al., 2019; Lu and Li, 2020; Wu et al., 2020; Lotfi et al., 2021). Although most research on information diffusion and rumor propagation contains text and content analysis, such features are location-specific and do not generalize universally (Siwakoti et al., 2021). Based on this, in our work, we employ a framework that consists of a temporal aspect in the form of an epidemiological model, and a graph-based structure in the form of the information cascade of each rumor.

Baseline rumor detection approaches are categorized based on their different feature engineering methods. The categories are (i) user characteristics, (ii) social context and content, and (iii) propagation structure characteristics. Handcrafted features based on user characteristics include such features as the number of followers, number of posts, and relevant information referring to the user of the social media platform (Kwon et al., 2013; Liu et al., 2015). Feature engineering employing social context, content, and linguistic attributes has been proven to yield important results. Such features include word sequences, phrase inquiries, and the relation of specific language with sentiment (Castillo et al., 2011; Yang et al., 2012; Kwon et al., 2013, 2017; Ma et al., 2015, 2016, 2018; Zhao et al., 2015; Liu and Wu, 2018; Yuan et al., 2020; Choi et al., 2021). Features based on the structural characteristics of the propagation have also been an integral part several rumor detection approaches (Ma et al., 2017, 2018; Liu and Wu, 2018). These graph-based features include information referring to the diffusion of a rumor in a social media network.

Several existing methods have explicitly proposed feature engineering, in order to improve the achieved accuracy (Castillo et al., 2011; Yang et al., 2012; Kwon et al., 2013, 2017; Liu and Wu, 2018; Ma et al., 2018). Below, we elaborate further on rumor detection approaches that include feature engineering and we discuss the baseline approaches of the field.

A decision tree classification (DTC) approach using word frequency, message, user, topic, and propagation characteristics to extract features has shown that such combinations allow for promising results (Castillo et al., 2011). Another content-oriented approach includes inquiring for specific phrases in the message (Zhao et al., 2015). The combination of content and user characteristics employing an RBF kernel Support Vector Machine (SVM-RBF) for classification (Yang et al., 2012) has demonstrated that the combination of such features is valid and achieves good results, while Support Vector Machine algorithms have also been used to classify features combining the temporal evolution of the content, along with the relevant social context and the underlying linguistically expressed sentiment with great success (Ma et al., 2015). A content-based approach using language pattern features from user comments and employing a Recurrent Neural Network (RNN) has also delivered important results (Ma et al., 2016), while others have attempted to combine the structure and content, in the form of semantic analysis, to perform the classification tasks (Ma et al., 2018). Along the same lines, content semantics and structural characteristics are the basis for feature engineering for similar attempts (Liu and Wu, 2018). More propagation-based heavy approaches have examined the temporal evolution of the propagation tree and classification takes place using an SVM classifier (Ma et al., 2017). A more holistic approach combines user characteristics, linguistic, and network features in a Random Forest classification scheme (Kwon et al., 2017).

In this section, we present the specific epidemiological models employed to model rumor propagation in online social media. We define and describe the function of the fitting parameters of these models, and we present the governing formulae for each model. Furthermore, we elaborate on feature engineering, regarding the graph modeling component of our approach, and we present the method that combines the two modeling practices in a rumor classification scheme.

In this section we present our experiments, including the necessary pre-processing, and our evaluation results. We present the results of rumor modeling and classification using epidemiological features, graph-based features as well as using their combination. Finally, we compare our method with other similar existing approaches to misinformation diffusion.

We propose a readily deployable solution for rumor detection on social media. The proposed framework is based on the diffusion cascade of an input rumor and does not require any additional user characteristics. This enables the use of the framework for any online social media platform, in contrast to existing literature that focuses only on Twitter. Furthermore, the framework does not require complex pre-trained language models or other high complexity content-based solutions and, thus, is applicable to real life systems, with no architectural adjustments or additional overhead.

Our method extracts temporal and structural features, enabling classification based on those features. The temporal-based features are extracted by fitting an epidemiological model on the original information diffusion data, and the structural features are gathered from a graph modeling of the same propagation data. We used the publicly available rumor propagation datasets Twitter15 and Twitter16 to evaluate the method. Experiments for binary (True, False) and 4-class (True, Non-Rumor, False, and Unverified) classification lead to the result that graph-based features perform better than epidemiological ones, while the most accurate classification is achieved through the combination of both feature sets.

Our method is location-independent and language-independent, leading to an approach that can be applied without consideration of the rumor content, in contrast to existing approaches that take into account rumor content. Our method is scalable and easily adaptable to existing real life systems with no modifications or additional overhead. However, it requires further investigation, due to dataset limitations. For example, our evaluations are for rumor cascades that begin from individual users, whereas in real life, multiple users may post the same rumor independently. Our method can accommodate multiple user sources, but no real dataset contains such information. Importantly, our method may provide improved results in that case, since machine learning techniques provide higher precision for large datasets.

The limitations of existing datasets are leading our future work, which focuses on collection of additional rumor propagation datasets from different online social networks and collection of data about rumors spreading from multiple root users, creating parallel diffusion trees. These datasets will enable us to evaluate the effectiveness of epidemiological models further and to evaluate graph-based features in rumor modeling and detection as well as overcoming some of the limitations of the Twitter15 and Twitter16 datasets.

Publicly available datasets were analyzed in this study. This data can be found here: https://www.dropbox.com/s/7ewzdrbelpmrnxu/rumdetect2017.zip?dl=0.

All authors contributed to manuscript revision, read, and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.