Machine learning (ML) is a branch of artificial intelligence (AI). Using algorithms learned via a large dataset, ML builds models with the ability to perform tasks without explicit learning, ranging from image classification to future prediction procedures [16] . Lying at the intersection of several scientific disciplines, ML is a rapidly growing field encompassing computer science, statistics, data science, and AI. ML has made significant strides with the advancement of algorithms and the abundance of data available through the Internet [17] . Rumor detection using ML frameworks is an exciting and rapidly developing research area. The ML process flow is depicted in Figure 2 , where the best characteristics are taken from the input and then categorized in a binary classification system as rumors (R) or non-rumors (NR).

In a supervised learning approach, models classify posts as rumors or not by training on a labeled dataset. This can be accomplished using several algorithms, including logistic regression, decision trees, the Naive Bayes algorithm, and support vector machines (SVM) [18] . Rumors can be accurately detected by filtering and analyzing language features using the supervised ML framework bifold [19] .

1) Methodology of Supervised Approaches

A substantial body of research on supervised approaches can be traced back to Castillo et al.’s [20] foundational study on assessing the reliability of information disseminated through social media networks examination of the reliability of information distributed through social media networks. Castillo et al. [20] , who were the first to develop features for evaluating credibility on Twitter [21] , based their research on the notion that users can assess the reliability of information using specific cues found in social media environments.

Castillo et al. created a dataset to investigate Twitter’s trustworthiness. A team of human assessors then assigned a label to each issue based on two criteria: NEWS, which provides a fact that may be of interest to others, and CHAT, which is a message based on the author’s personal ideas and/or chats with friends. Another panel of judges rated each item’s trustworthiness within these categories [21] . Subsequently, a collection of attributes was extracted to construct a classifier that automatically ascertains whether a subject is associated with noteworthy data or occurrences and, if so, automatically evaluates the subject’s degree of believability. Four categories of features were distinguished: propagation-based, message-based, user-based, and topic-based features [21] .

The best-first selection approach was used to select features that contributed most to the task. Sampling with replacement was used to select a random sample from the training dataset, and a uniform distribution was used to determine the likelihood of obtaining an instance from each of the three classes (NEWS, CHAT, and UNSURE). A variety of supervised learning techniques were used [21] .

2) Nonparametric Machine Learning Models

The term nonparametric refers to models in which the underlying data distribution is not assumed to have a fixed shape or a fixed number of parameters. Instead, they adjust to the structure of the data, which can be used with both supervised and unsupervised ML models. In classifying fake news according to its prevalence level, in addition to the use of textual information (text and headlines) and content features, preventive strategies can also be implemented to create early predictions regarding the spread of fake news [22] .

3) Logistic Regression Model

Logistic regression models are statistical models used to estimate the probability of a binary event, such as yes/no, true/false, and, in our context, common/uncommon, using one or more independent variables [23] . Logistic regression has been used in the health sector to classify health-related rumors. According to a study on rumor veracity categorization, the length of rumor headlines or statements and the use of pictures in rumor statements were adversely connected with the chance that a rumor was real. The following characteristics were positively connected with the likelihood that a rumor would be true: hyperlinks, the names of people or places mentioned in the rumor, and hints about the source of information [24] .

Using regression tasks, fact-checking platforms are developing computational methods to predict the popularity of false rumors, focusing on immediate contextual information, such as rumor content and user profile information [25] .

4) Support Vector Machines (SVMs)

In the 1950s, the SVM method first appeared, subsequently developing as a kernel computer [16] . The SVM is a supervised ML model that is applied to classification and regression problems and works by ideally dividing data points into different classes to determine the ideal hyperplane in the feature space [26] .

In this study, Qin et al. used an SVM on a training set and based rumor prediction on a fixed threshold technique to maximize prediction accuracy. Furthermore, they used pseudo-feedback, novelty-based, and content-based features. The proposed method can reliably forecast whether an item eventually becomes a rumor, enabling the refutation of false information before it spreads and causes harm to society. Rumor prediction significantly increases the precision of cutting-edge rumor detection software [5] .

5) Random Forest (RF)

The RF model, which is also widely used for categorization, uses several decision trees. In this model, each decision tree serves as a classifier, and a random sampling algorithm aggregates all classification results before the final classification decision is reached. RF has been demonstrated to be a highly accurate and reliable algorithm compared with various rumor-refuting microblogs [27] . However, it has not been proven to be as effective as XGBoost in identifying users who are more likely to deny rumors [28] .

6) Extreme Gradient Boosting (XGBoost)

XGBoost is a recently developed algorithm that is increasingly popular in data analysis and industrial circles. In contrast to RF, it trains trees in a serial and interactive manner, as opposed to a parallel and autonomous manner, using a different type of tree called the classification and regression tree (CART) [29] .

XGBoost was used in a study to determine which users were more inclined to deny rumors. The prediction model was enhanced by text analysis using natural language processing (NLP), which also demonstrated high generalizability and the absence of any significant group behavior characteristics in the retweeted prediction. A valid and robust distinguishing XGBoost model was then trained and validated to predict who would retweet disaster-related rumor-refuting microblogs [28] .

The XGBoost model has been used, along with several content-, user-, and topic-based features, to automatically detect rumors in Arabic [30] . In another study, an XGBoost classification model was trained utilizing NLP, together with other user characteristics, such as user sentiment and current interests, allowing it to examine the relationship between user characteristics and rumor-disputing behavior in five major rumor categories—economics, society, politics, disasters, and the military [31] .

The unsupervised learning approach involves training models on unlabeled data with the goal of uncovering underlying structures or hidden patterns in the data. Illustrative techniques encompass anomaly detection and dimensionality reduction (principal component analysis, t-small number ensemble), as well as clustering (K-means, linear clustering) [16] [32] .

1) Feature Analysis

Feature analysis is an unsupervised learning method that aims to identify hidden patterns or structures in data. In our context, many features have been examined for rumor detection, such as retweet rate and word distribution characteristics. A study concluded that the retweet rate was insufficient to distinguish rumors from other items. In contrast, the differences in word distribution may be an important indicator for classifying rumors [33] .

2) Clustering

Clustering is a method that uses models to group data points based on the similarity of the data. It seeks to identify groups in a dataset that occur naturally without labels. Examples of clustering algorithms include K-means, hierarchical clustering, and DBSCAN [34] .

A study on clustering Twitter political rumor detection used five structural and chronological characteristics. In the first phase, tweets with consistent address links were grouped. In contrast, in phase two, similar clusters debating continuous news were combined into a single cluster with similar circular functions [35] . Using clustering to group similar tweets and semantic and sentiment analyses to identify differences in information supplied by various sources, these techniques work together to help uncover rumors on Twitter [36] .

This method uses unlabeled data by generating supplementary tasks to enhance the primary supervised task. To better identify and detect rumors, the self-supervised rumor detection (SRD) framework employs contrastive learning on heterogeneous information sources and is part of the unsupervised learning subcategory. This technique does not rely on labels from other sources but instead on the data itself to generate supervisory signals. The applications used in this framework include NLP, computer vision, and speech recognition. Common techniques include removing the signal from the image, which includes filling in the missing parts of the image, contrastive learning, and word prediction using language models such as BERT (bidirectional encoder representations from transformers) [37] .

1) Contrastive Learning

By improving the model’s capacity to discern between accurate and inaccurate information, contrastive learning has the potential to be a valuable tool for identifying rumors. It has been used in conjunction with other technologies, which enhanced performance [38] .

2) BERT Models

Researchers have demonstrated that using BERT in different ways significantly enhances the accuracy of rumor detection compared with traditional methods [39] [40] . Such increases are owing to BERT’s enhanced capacity to comprehend the subtleties and contextual meaning of language used in rumors.

Deep learning is a subfield of ML that employs multilayered artificial neural networks (ANNs), popularly known as “deep” neural networks (DNNs), to represent complex patterns in data. Neural networks can acquire knowledge from vast datasets because of their inherent design, which emulates the cognitive processes of the human brain. Deep learning works well for applications such as audio and picture recognition and NLP [41] [42] and has demonstrated to be a clear winner over traditional classifiers in a variety of ML applications, including speech recognition, object identification, and sentiment categorization. The goal of DNN-based techniques is to automatically develop deep representations of rumor data [43] . The use of deep learning for rumor detection is illustrated in Figure 3 .

Several deep learning methods are employed for rumor detection on social media. The subsections below outline some common approaches.

A diverse class of feedforward neural networks, RNNs are mostly employed to handle variable-length sequential or time series data. RNNs are used in a variety of applications, such as categorization and sequence creation [3] . Given their structure, RNNs form models of rumor data as sequential data, capturing the dynamic temporal signals characteristic of rumor diffusion through direct cycle connections between units [43] . For instance, integrating an RNN model with a Word2Vec model provides a novel approach to rumor detection. This hybrid model comprises two primary components: Word2Vec and long short-term memory (LSTM)-RNN. A common problem faced in detecting rumors in breaking news is overlapping topics and out-of-vocabulary (OOV) items; parallel training of components helps mitigate such issues, providing a powerful solution for detecting rumors in real time. An RNN-based approach was used to leverage both the propagation patterns and linguistic content of posts for early rumor detection, allowing the RNN to learn the relationships between posts and their content [44] .

The CSI model, which refers to the components that make up the model: Capture, Score, and Integrate, is an innovative rumor detection technique that combines text, user comments, and promotions. The use of an RNN enables the understanding of the dynamics involved in the spread of rumors. The hybrid CSI model enhances accuracy and reliability by integrating several data profiles. Furthermore, it offers valuable insights for further studies by generating concealed user and object representation. The CSI model represents a significant advancement in rumor detection. Providing a comprehensive reaction to misinformation shared on social media enables swift intervention and mitigation [45] .

A CNN utilizes a structure that models significant semantic features, thereby enabling it to extract local and global features and reveal high-level interactions [43] . Recent research has revealed constraints in using RNNs for the early identification of rumors because they rely heavily on recent input data. A solution to this problem was the development of a convolutional approach for misinformation identification (CAMI). The proposed approach uses doc2vec to represent clusters of microblog entries and a two-layer CNN for classification, enabling the efficient extraction of crucial information and high performance [46] .

In another study, a novel event adversarial neural network (EANN) was developed specifically for detecting multimodal false news, particularly for recent events. The EANN incorporates modules for event discrimination, bogus news identification, and multimodal feature extraction. It incorporates CNNs to extract characteristics from both textual and visual data [47] .

These neural networks are designed to operate with graph-structured data. GNNs are especially helpful for tasks in which relationships and interactions between entities are crucial. Among its most important applications, social network analysis is used to detect communities, predict user behavior, and identify influential users, with other applications including recommendation systems and NLP. The global structural characteristics from graphs or trees are better captured by a graph convolutional network (GCN) than by the aforementioned deep learning models. Driven by CNN’s achievements of CNNs in computer vision, GCN have proven their performance at the forefront of several tasks using graph data. Several recent studies have used GCNs to detect and predict rumors on social media networks [48] .

Deep learning methods often overlook the structures of wide dispersions in rumor detection. The bi-directional graph convolutional network (Bi-GCN) model explores both the propagation and dispersion characteristics of rumors using a top-down directed graph and an opposite graph of rumor diffusion. The proposed model has been proven to be effective in early rumor detection on Twitter [49] .

A gated GNN-based algorithm [50] called propagation graph neural network (PGNN) can generate powerful representations for each node in the graph. The global embedding with PGNN (GLO-PGNN) and ensemble learning with PGNN (ENS-PGNN) models adopt different classification strategies and include attention mechanisms to dynamically adjust the node weights to further improve performance.

The use of a unified GNN model and a community-enhanced vulnerability propagation method can improve rumor prediction and identify user vulnerabilities [51] . A gated GNN was used to learn semantic dependencies across segments in a graph-based pivotal semantic mining framework [52] . Given the many shortcomings of previous techniques and the need for improved reliability and accuracy of detection mechanisms, these sophisticated models and methodologies mark a substantial advancement in the field of rumor and false news identification.

Hybrid models combine more than one approach to increase the accuracy and effectiveness of the model. In the context of detecting rumors, hybridization often combines traditional ML methods with deep learning to enhance the effectiveness of the model and its ability to detect rumors [4] .

In rumor detection, hybrid methods can be classified as follows:

This model combines several features that enable it to predict rumors with higher accuracy. The combination of features such as text, user behavior, and network structure increase the model’s effectiveness and accuracy [38] [40] [53] [54] .

The ensemble learning approach can combine the capabilities of more than one model in prediction to improve and increase efficiency, which makes it superior to single models in terms of their accuracy and flexibility. Examples of common methods for this approach include boosting, stacking, and bagging [38] .

Multimodal approaches employ several features together to offer a comprehensive view of a situation. In predictions, these approaches can use a combination of text, images, and audio to enhance model functioning [55] - [57] .

More than one algorithm can be exploited to explore the benefits that these methods provide since most ML algorithms, and many neural network structures can be complemented with deep learning methods for higher levels of accuracy and enhanced analysis [39] [54] [56] [58] [59] .

This method uses several deep learning techniques, like CNNs and RNNs, to create a more accurate and useful model or way of predicting or extracting features that make use of the strengths and weaknesses of many neural network designs [38] [40] [55] [56] [59] .

For rumor detection, network-based approaches have been proposed that focus on the structure of social networks to identify how information and rumors are distributed and spread across them. Community discovery algorithms, centrality measures, and influence modeling often aid in the identification of influential nodes or groups within a community, which is a major factor in the spread of rumors across the network. By tracking the information spread and rumor pathways in hypothetical networks, network-based approaches expect to forecast, observe, and eventually control the harm and negative consequences of rumors [21] .

The reliability analysis of rumor propagation modeling is currently one of the most well-known multidisciplinary study areas regarding this topic. Existing studies have primarily focused on three areas:

The enhancement of the classical rumor propagation model based on factors, such as, heterogeneous user representation, was proposed for automatically distinguishing rumors from credible messages posted on social media [21] . The analysis of rumor propagation on online social platforms through the integration of other disciplinary theories, such as game theory, where social anxiety is considered, the intensity of rumors, and public perceptions of rumors and anti-rumors are considered important factors [60] . The combination of more than one technology that focuses on prominent nodes to reduce the propagation of rumors in online social networks has been shown to increase the effectiveness of control mechanisms when high-centrality users are the focus of attention [61] .

Social networks enable individuals to participate in the exchange of information, providing large amounts of diverse data. It is worth noting that conventional techniques for social circle finding, similar to community detection, use clustering through statistical inference from a network topology perspective. In contrast, global mining requires considerable resources due to the size of social networks. Because of this difficulty, many academics have concentrated on mining the social circles of users’ egocentric networks. Most social circle-finding algorithms are only applicable to specific datasets, making it difficult to work directly with them to detect rumors [6] [62] .

Following the remarkable success of transformer-based embeddings in natural language tasks, researchers have become interested in applying them to classify rumors on social media platforms. Unlike traditional word embedding methods, the former excels at capturing the contextual meaning of a text by examining words to the left and right, producing textual representations that are ideal for detecting rumors [63] . A novel approach, called the ‘Memory-Enhanced Transformer with Graph Convolutional Networks’ (GCNs-MT), was presented for detecting rumors on social media platforms [64] . By combining GCNs with long-short-term memory cells and multi-head attention mechanisms in transformers to capture local and global dependencies in rumor propagation, we have proven our ability and efficiency in detecting rumors, addressing the challenges posed by the rapid spread of rumors on social media platforms.

As fake news and rumors spread rapidly, it is important to develop a model for automatic rumor detection. The transformer-based multimodal interactive fusion (TMIF) model [65] relies on a transformer to map multimodal feature representations to the same data domain for fusion purposes. It can capture multilevel dependencies between multimodal content while minimizing the impact of multimodal heterogeneity. Furthermore, it demonstrated superior performance in the early detection of rumors.

The examination of rumors on SM has become a major area of interest in SM analysis. Several researchers have examined the way in which unverified information moves in these digital spaces with the sole aim of understanding these dynamics and their effects. Such research has made significant contributions to the study of rumor propagation and has benefited our understanding of how false information moves in online spaces. Several studies have also contributed to describing the various dimensions in which rumors operate in the quest to explain the inter- or intraspecific relationships between human behaviors, network structures, and how information flows in these digital ecosystems.

The dynamics of rumor propagation have been investigated in multiple studies that have sought to reveal the strands of a rumor on the paths that people follow on SM networks. Many of these studies have used computational models, network analyses, and empirical studies to scrutinize how rumors start, gain, and lose momentum, spread across diversified user groups, and sometimes go viral. Most of these studies have employed network-based approaches and empirical analyses to study the patterns of rumor dissemination on X/Twitter. Network-based approaches, such as social network analysis, have identified influential users and network structures that provide avenues through which rumors can have longer lifespans [6] .

Platforms such as Twitter/X and Reddit provide publicly available data, enabling researchers to more effectively study and track rumor spread patterns, user interactions, and content characteristics due to their open and accessible nature. In contrast, other social media platforms such as WhatsApp and Snapchat are important for rumor detection research due to their widespread use. Rumors on these platforms often spread through direct user interactions, making tracking difficult due to their closed nature, which limits the availability of public datasets, which are critical for rumor detection research. Researchers are therefore exploring innovative algorithms such as metadata analysis or leveraging trust mechanisms to address such challenges [69] [70] .

Studies have demonstrated the varied characteristics of rumors about SM. For instance, rumors can go viral, which means they spread quickly and widely through SM platforms because users are all connected to each other [71] . Rumors can take advantage of the fact that the truth about a story is not known, making them more ambiguous and allowing individuals to form assumptions about the truth. Moreover, rumors are strengthened by individual beliefs, fears, or desires when shared through social circles. Rumors are also interesting because they change as they travel from friend to friend, or in our case, platform to platform; thus, they are harder to track and control [71] . To account for this semantic drift in rumors, models require innovative techniques. For example, using temporal segmentations instead of random segmentations to train models on past data and test them on future data [72] . Dynamic embedding models, such as those based on word2vec or BERT, are also a good method for monitoring changes in word meanings and context over time [39] . Analyzing the path of rumor propagation by incorporating contextual implications of propagation paths provides a deeper understanding of rumor propagation, both structurally and semantically. Combining these two factors enhances the model’s ability to adapt to the changing nature of rumors [61] .

Furthermore, the spread of rumors is not solely reliant on their content; the structural properties of the underlying social networks also play a crucial role. Some methodologies have emphasized the need to debunk rumors in real time on social networks using the hybrid clustered shuffled frog-leaping algorithm with particle swarm optimization (SFLA-PSO), as well as taking advantage of trust relationships [58] .

The present study, which examined friendship maintenance in social networks, highlighted the importance of trust in these digital relationships. The spread of rumors often threatens the dynamics of social networks; therefore, understanding how rumor propagation interferes with these dynamics is crucial [58] [73] .

Both [58] and [73] made substantial contributions to the field by exploring the creation of a comprehensive trust mechanism. Their research highlights the significance of trust levels in online social network (OSN)-based social relationships, emphasizing this aspect over intimacy levels. Rumor detection in social networks can be a complex task because the state of information verification can change over time. Several recent studies have explored temporal classification to address the dynamic nature of rumors [74] . For example, the rumor detection method used in this study constructs a distributed spreading tree where the weight of each edge represents the temporal interval between connected posts. This method incorporates temporal information to model the spread of rumors over time, which improves the accuracy of rumor prediction and reduces the din [75] .

As illustrated in Figure 4 , at the rumor generation stage, a rumor frequently goes through numerous iterations of content within the social circle before making its way across social networks.

Social circle mining is an analytical system that helps us understand the structure of social networks on digital platforms, such as online communities, and their functioning. SCM is used to extract, interpret, and analyze social circles from networks and employs various methodologies, including network science, ML, and data mining, to uncover communities and relationships within large networks [76] . The primary purpose of SCM is to identify and characterize clusters or groups of individuals with strong interconnections or shared interests within a wider social network [77] [78] . SCM uses methods including the traditional K-means clustering algorithm [79] and hierarchical clustering [80] , as well as advanced algorithms such as community discovery algorithms based on module optimization or spectral clustering. Such methods uncover the structure of social networks, which leads to a deeper understanding of how they are organized [81] [82] .

A practical example of this is the use of social circle mining [6] , which extracts highly homogeneous social circles from user context, and combines them with social interaction to detect rumors. This method, called RDMSC, is processed and used in GCN encoders with an attention mechanism to learn and combine the two types of features for classification, as shown in Figure 5 . The results demonstrate that the RDMSC method achieved superior rumor detection capabilities of up to 95%, outperforming state-of-the-art baselines on three real-world datasets.

Consider a practical example where SCM is integrated with a deep learning model for rumor detection. Data is collected from tweets related to a trending topic. The data is cleaned by removing retweets, filtering out non-English tweets, and normalizing text data. Hierarchical clustering is used to identify groups of users who frequently interact with each other. Features such as the number of interactions within each circle, sentiment of tweets, and frequency of keyword usage are extracted. These features are integrated into an LSTM (Long Short-Term Memory) network to capture temporal dependencies and patterns in the data. The LSTM network is trained to classify tweets as rumors or non-rumors based on the extracted features. The model’s performance is evaluated using a validation dataset, and it is refined based on the results. Such examinations enable the study of various social media research problems, such as echo chambers or filter bubbles, how information is diffused, the important factors contributing to information becoming viral in each social circle, and the trajectory of information diffusion (such as rumor debunking) [83] [84] . Moreover, SCM also helps deliver targeted content because, through SCM, a platform can ascertain social circles based on user interaction and shared interests, such as movies, stock market information, or religious factors, enabling platforms to prompt collective behavior by targeting groups with those shared interests [82] [85] . Furthermore, SCM holds significant promise for rumor detection and mitigation. SCM is capable of understanding how rumors spread within a particular community by analyzing the structural properties of social circles, such as identifying influential nodes in a diffusion process and mapping diffusion paths [6] [84] . Furthermore, this capacity can facilitate the development of proactive strategies to intervene and counteract the spread of misinformation. It can also help mitigate its impact by targeting influential nodes and disrupting the pathways of rumor diffusion within these social circles.