User:Tazin007/Sample page

Review Paper
= Deep learning for misinformation detection on online social networks: a survey and new perspectives =


 * Md Rafiqul Islam1 · Shaowu Liu1 · Xianzhi Wang 2 · Guandong Xu1 Received: 28 March 2020 / Revised: 11 September 2020 / Accepted: 12 September 2020 / Published online: 29 September 2020 © Springer-Verlag GmbH Austria, part of Springer Nature 2020

Abstract
In recent times, the widespread use of social media platforms like Facebook, Twitter, and Sina Weibo has become an integral part of our daily lives, offering users a convenient means to share personal messages, photos, and videos. However, the ubiquity of social networks has also given rise to deceptive practices, such as the spread of fake news and rumors, which can lead users to believe in false information. This proliferation of misinformation on social networks has evolved into a global concern, leading to a surge in research interest in the field of misinformation detection (MID). Numerous studies have emerged, addressing new research challenges and techniques in this area. Detecting misinformation automatically presents a significant challenge as it requires advanced models to discern the relevance of reported information compared to factual data. Previous research has primarily focused on categorizing misinformation into three main types: false information, fake news, and rumors. In response to these issues, this survey offers a comprehensive overview of automated misinformation detection encompassing false information, rumors, spam, fake news, and disinformation. It underscores the pivotal role of deep learning (DL) in processing data, extracting patterns, and making informed decisions, thereby enhancing the extraction of global features and achieving superior results. The survey underscores DL's effectiveness and scalability in state-of-the-art MID. It concludes by highlighting existing limitations that hinder real-world implementation and pointing towards future directions for further exploration in this domain.

1 Introduction
Online social networks like Facebook, Twitter, and Sina Weibo have become an integral part of our daily lives, enabling people to share their thoughts, videos, and news about their activities (Gao and Liu 2014; Islam et al. 2018a). While these platforms offer a convenient way for users to connect, they also provide fertile ground for deceptive activities, such as the spread of fake news and rumors, which can mislead users (Kumar et al. 2016). Consequently, the field of misinformation detection (MID) in social networks has garnered significant attention as a burgeoning area of research (Wu et al. 2019; Goswami and Kumar 2016). Detecting misinformation in an automated manner is challenging, as it necessitates advanced models to assess the relationship between reported information and real data (Wu et al. 2019). In response to these challenges, both academia and industry researchers have turned to deep learning (DL) techniques to make informed decisions in addressing various complex MID problems (Xu et al. 2019; Yenala et al. 2018; Yin et al. 2020). Therefore, this survey aims to provide a systematic overview of the current state of research in MID, with a particular focus on DL techniques.

Social networking sites have evolved into dynamic platforms used for diverse purposes, including education, business, healthcare, and telemarketing, but they have also become susceptible to illicit activities (Vartapetiance and Gillam 2014; Wu et al. 2019; Naseem et al. 2020). People primarily utilize these networks to interact with friends and colleagues, engage with customers, and glean valuable insights for business trends (Bindu et al. 2017). The exchange of information happens rapidly, irrespective of geographical boundaries. For example, businesses use social media for marketing and adjusting their product offerings based on market demand and customer feedback. Celebrities leverage these platforms to enhance their public presence, and government entities employ them to collect public opinions. However, the openness of social networks also paves the way for harmful activities, where incorrect or misleading information is spread (Bharti et al. 2017; Sun et al. 2018; Gao and Liu 2014). Consequently, the field of social network analysis has grown into an interdisciplinary research area aimed at exploring and adapting techniques to analyze global social network data. While previous studies may have approached the concept of misinformation differently, this survey emphasizes the timeliness of misinformation detection in social media (Sharma et al. 2019; Shu et al. 2020).

Misinformation refers to inaccurate information created to deceive readers (Fernandez and Alani 2018; Zhang et al. 2018a). This category includes various forms of misinformation such as fake news, rumors, spam, and disinformation, which typically encompass numerical, categorical, textual, and image data, and often result in detrimental consequences (Ma et al. 2016; Bharti et al. 2017; Helmstetter and Paulheim 2018). Due to the extensive use of social media, unscrupulous individuals have opportunities to disseminate misinformation through fake accounts, often presenting their information convincingly with proper references. However, identifying and categorizing misinformation, including fake reviews, false information, and rumors, is essential to thwart its impact. Detecting misinformation in social network data can offer early insights into emerging issues like stock market fluctuations, political rumors, social matters, and business performance (Habib et al. 2019). Over the years, various techniques have been employed to distinguish genuine from fraudulent information or users (Islam et al. 2018a; De Choudhury et al. 2013a; De Choudhury et al. 2013b). However, traditional methods struggle to address the myriad forms of misinformation effectively, making deep learning-based detection approaches suitable for accommodating a wide range of features for MID.

The growth of machine learning (ML) and deep learning (DL) techniques has garnered significant attention in both industry and research communities (LeCun et al. 2015; Islam et al. 2019). Notably, DL-based approaches have become a cornerstone of MID. A substantial body of research has explored automatic MID as well as related areas, including rumor detection, fake information, and spam detection. This survey aims to comprehensively review the utilization of DL techniques in MID, elucidating the specific challenges and solutions. Although earlier surveys have examined MID, we find that there is no clear distinction between different forms of misinformation, and there's a lack of systematic DL-based reviews across various types of misinformation problems. The ever-evolving nature of misinformation, fueled by advances in large-scale pre-trained models and adversarial learning, calls for the use of high-capacity models such as DL. Therefore, our focus is on MID, offering detailed discussions of DL techniques and their limitations.

While existing surveys have covered a broad spectrum of techniques for MID, our survey focuses on DL methods' increasing role in detecting misinformation. We provide insight into how various MID problems align with different DL techniques, which has not been extensively covered in previous surveys. We also highlight ongoing issues that warrant attention, such as data volume, data quality, explainability, domain complexity, interpretability, feature enrichment, model privacy, incorporating expert knowledge, and temporal modeling. In summary, our survey's main contributions include:


 * 1) A comprehensive systematic review of existing problems, solutions, and validation in MID within online social networks using various DL techniques.
 * 2) An analysis of the strengths and limitations of various techniques, emphasizing the role of DL in addressing misinformation on social networks.
 * 3) A discussion of open issues and promising directions for future research.

The remainder of this paper delves into the formal definition of MID, its types, impacts, and challenges. It then explores state-of-the-art DL techniques for MID and outlines future research directions and open issues. Finally, we conclude the survey.

2 Background
In this section, we focus on the theme of misinformation detection (MID) and delve into its formal definition, various types, and the impact it has on social networks (SN). However, it's important to note that this section does not delve into the techniques for MID or their effectiveness. The introduction of this paper has highlighted the emergence of deep learning (DL) as a critical player in MID. The primary objective of this paper is to review the role of DL in the process of MID. In the upcoming sections, we will discuss the significance of DL in MID, provide an overview of its current performance, and outline some potential future directions.

2.1.1 What is misinformation?

Misinformation is essentially the dissemination of false statements intended to mislead people by concealing the correct facts. It can take on various forms such as deception, ambiguity, and falsehoods. Misinformation can generate feelings of mistrust, leading to weakened relationships, which is essentially a negative breach of trust. People generally expect honest and truthful communication, whether it's from close friends, relatives, or strangers. As an example, consider a Facebook discussion about a recently released product where both genuine users and fake users participate. While real users discuss the product's features honestly, fake users offer praise for the product regardless of their actual opinion.

To illustrate the problem, let's say you have a collection of restaurant reviews from a set of users, among which are both genuine feedback and fabricated reviews created by restaurant owners to deceive. Distinguishing these false reviews from authentic ones can be a challenging task, and the role of the researcher is to identify and differentiate the real from the fake.

2.1.2 Types of misinformation

There are several terms related to misinformation, including rumor, fake news, false information, spam, and disinformation. Rumors are essentially unverified stories that circulate from person to person, often with uncertain truthfulness. Fake news refers to intentionally misleading news articles that can be verifiably proven false. Misinformation, as a broader term, is used to describe information that is untrue. Spam, on the other hand, involves sending unsolicited messages over the internet, typically for malicious purposes like spreading malware or advertising. Disinformation is a subset of misinformation, but it's characterized by the deliberate spread of false or misleading information. The key distinction between misinformation and disinformation lies in intent, with misinformation often spread without the intent to deceive, whereas disinformation is spread with the specific aim of misleading others.

Numerous studies have been conducted on misinformation identification in social media, and various approaches have been developed. Some methods treat individual microblog posts as objects, assess their credibility, and then aggregate these assessments at the event level. Others extract various features at the event level to determine whether an event involves misinformation. Different studies also employ hand-crafted features, including conflict viewpoints, temporal properties, user feedback, and signals from tweets containing skepticism. This section defines and describes each type of misinformation in detail.

2.1.3 Impact of misinformation

Misinformation can have far-reaching consequences across various aspects of life, affecting the social, political, economic, stock market, emergency response, and crisis events. It is often used to manipulate public opinion, influence political elections, and even pose threats to public security and social stability. In many cases, misinformation revolves around fabricated information related to fictional issues rather than providing relevant or accurate information. With the advent of social network platforms like Facebook, Twitter, and Sina Weibo, the rapid spread of misinformation has become more prevalent. In conversations, people can share information presented as factual, even when it may not be entirely true. This misleading information is sometimes shared for personal gain, especially in the context of political issues. Misguided residents, who believe and express this misinformation with certainty, can significantly influence elections.

The rise of social networks and technological advances has led to an explosion of misinformation, with numerous studies conducted to measure its impact. For instance, studies have examined the spread of rumors on Facebook and the use of fake news articles during elections. The negative impact of misinformation is evident, such as how organizations can undermine reliable evidence through deceptive campaigns. In healthcare, misinformation can lead to dangerous outcomes, as seen with public skepticism regarding vaccines.

While existing research on misinformation often focuses on text content, with some exceptions for image and video content, the effectiveness of different MID approaches remains a challenge. The challenges in MID encompass data volume, data quality, domain complexity, interpretability, feature enrichment, model privacy, incorporation of expert knowledge, temporal modeling, and dynamic factors. To address these challenges, the paper introduces deep learning (DL) as an emerging state-of-the-art technique for MID.

2.2 Deep Learning

Deep learning (DL) is a term that was initially introduced to the machine learning community by Dechter in 1986 and to artificial neural networks based on a Boolean threshold by Aizenberg in 1999. DL has emerged as a powerful technique in various domains of artificial intelligence, such as computer vision, speech recognition, natural language processing, anomaly detection, and more. It has the potential to enhance learning processes, expand the scope of research, and simplify measurement procedures.

In recent years, researchers have explored the application of DL to solve a wide range of problems in online social networks. DL has gained popularity due to its potential in various domains, and researchers have been using different neural network architectures, including convolutional neural networks (CNN), recurrent neural networks (RNN), and long short-term memory (LSTM), to explore their application in different contexts.

DL is capable of handling complex tasks and is gaining traction globally. While it may require more time for data training, it offers quick testing and is increasingly used across various domains. To expedite DL processing, DL frameworks have been developed to simplify the implementation of modularized DL algorithms, optimization techniques, and distribution techniques. These frameworks are designed to facilitate system-level development and research. Some popular DL frameworks, like Caffe, Torch, TensorFlow, MXNet, and CNTK, provide researchers with the tools they need and offer a higher level of abstraction to simplify complex programming challenges. Most of these frameworks are implemented in Python, a widely used language for DL architecture design, making programming more efficient and accessible.

3 Deep learning for misinformation detection: state-of-the-art
Misinformation detection is defined as an observation that deviates greatly from other observations and thereby arouses suspicion that it was generated by a different mechanism. In Sect. 2.1.2, we have discussed different terms related to misinformation with examples. It is observed that the same type of problem has been solved by many techniques. Although many techniques are being used to detect misinformation in social network data, DL is one of the better approaches to use. However, the same type of misinformation problems has been solved with various DL techniques (Table 3). Additionally, these types of DL techniques are dependent on different data characteristics and used to automatically identify misinformation. Therefore, we have divided the DL techniques into three main categories based on the model as follows: (1) discriminative models, (2) generative models, and (3) hybrid models. All three categories have a large number of architectural models that are commonly used for MID. However, due to differences in performance, we only discussed 12 models namely convolutional neural networks (CNN), recurrent neural networks (RNN), recursive neural networks (RvNN), restricted Boltzmann machines (RBM), deep Boltzmann machines (DBM), deep belief networks (DBN), variational autoencoders (VAE), convolutional restricted Boltzmann machines (CRBM), convolutional recurrent neural networks (CRNN), ensemble-based fusion (EBF), and long short-term memory (LSTM), as shown in Fig. 4. We discuss each model that uses for MID, respectively.

3.1 Discriminative model for detecting misinformation
A variety of discriminative models used social content and context-based features for MID. In recent years, to tackle the problem of misinformation, several studies have been conducted and revealed some promising preliminary results. Therefore, we briefly review the three discriminative models namely CNN, RNN, and RvNN, respectively. It is noted that the discriminative-based models have demonstrated significant advances in text classification and analysis.

Convolutional Neural Network (CNN) CNN is one of the most popular and widely used models for the state-of-the-art of many computer vision tasks (LeCun et al. 2010). However, recently, it has been extensively applied in the NLP community as well (Jacovi et al. 2018). For example, Chen et al. (2017) introduced a convolutional neural network-based classification method with single and multi-word embedding for identifying both rumor and stance tweets. Kumar et al. (2019) introduced both a CNN and a bidirectional LSTM ensembled network with an attention mechanism to solve MID. Additionally, Yang et al. (2018) stated that online social media is continually growing in popularity and genuine users are being attacked by many fraudulent users. They informed that fake news is written to intentionally mislead users. In their paper, they applied the TI-CNN model to identify the explicit and latent features from the text and image information. They demonstrated that their model solves the fake news detection problem effectively.

Recurrent Neural Network (RNN) RNN utilizes the sequential information in the network which is essential in many applications where the embedded structure in the data sequence conveys useful knowledge (Alkhodair et al. 2020). The advantage of RNN is its ability to better capture contextual information. To detect rumors, existing methods rely on handcrafted features to employ machine learning algorithms that require a huge manual effort. To guard against this issue, the earliest adoption of RNNs for rumor detection is reported in Ma et al. (2016) and recurrent neural networks with attention mechanism in Chen et al. (2018) and Jin et al. (2017b). Figure 5 shows the RNN architecture used for the fake news detection proposed by (Shu et al. 2019a). Authors have proposed different RNN architectures, namely tanh-RNN, LSTM and Gated Recurrent Unit (GRU) (Cho et al. 2014a). Among the proposed architectures, GRU has obtained the best results in both the datasets considered, with 0.88 and 0.91 accuracy, respectively. Ma et al. (2016) proposed a RNN model to learn and that captures variations in relevant information in posts over time. Additionally, they described that RNN utilizes the sequential information in the network where the embedded structure in the data sequence conveys useful knowledge. They demonstrated that their proposed model can capture more data from hidden layers which give better results than the other models.

Recursive Neural Network (RvNN) Researchers are more concerned to identify unscrupulous users in SN and want to protect genuine users from fraudulent behavior (Guo et al. 2019). Therefore, RvNN is one of the most widely used and successful networks for many natural language processing (NLP) tasks (Socher et al. 2013; Zubiaga et al. 2016a). This architecture processes objects that can make predictions in a hierarchical structure and classifies the outputs using compositional vectors. To reproduce the patterns of the input layer to the output layer, this network is trained by auto-association. Also, this model analyzes a text word by word and stores the semantics of all the previous texts in a fixed-sized hidden layer (Cho et al. 2014b). For instance, Zubiaga et al. (2016b) proposed a RvNN architecture for handling the input of different modalities. Ma et al. (2018) proposed a model that collects tweets from Twitter and extracts features from discriminating information. It follows a non-sequential pattern to present a more robust identification of the various types of rumour-related content structures.

3.2 Generative model for detecting misinformation
In recent decades, online social media platforms have become fertile grounds for the proliferation of deceptive opinions, including rumors, spam, trolling, and fake news, all of which are deliberately crafted to appear authentic. Many existing approaches to Misinformation Identification (MID) rely on syntactic and lexical patterns or opinion-related features. This section explores the application of five generative models in various MID applications:

Restricted Boltzmann Machine (RBM): RBM is a generative stochastic artificial neural network capable of learning a probability distribution over its set of inputs. It efficiently trains multiple layers of hidden units without intra-layer connections between them. While RBMs have been used in various applications, their application in MID is relatively limited, with recent attempts focusing on using RBMs for feature extraction related to spam detection.

Deep Belief Network (DBN): DBN is a generative graphical model with multiple layers of latent variables or hidden units. It is constructed from simpler unsupervised networks like RBMs. DBNs have seen wide adoption across various domains, including spam detection and the identification of false data injection attacks in smart grids. Research indicates that DBN-based methods outperform traditional approaches.

Deep Boltzmann Machine (DBM): DBM is a binary pair-wise Markov random field with multiple layers of hidden random variables. It has been applied to detect malicious activities, especially in the context of fake news detection. A multimodal deep learning model based on DBMs has been employed to improve the accuracy of fake news identification.

Generative Adversarial Network (GAN): GANs are a class of machine learning systems that learn to generate new data with statistics similar to a given training set. They have been used to enhance automated rumor detection, making it more robust and efficient. GANs identify features related to uncertain or conflicting voice production in rumors.

Variational Autoencoder (VAE): VAE models make strong assumptions about the distribution of latent variables and use a variational approach for latent representation learning. VAEs have played a vital role in extracting new patterns by analyzing user responses to true and false news articles, contributing significantly to the detection of misinformation on social media platforms.

These generative models offer innovative and promising solutions to address the challenges of misinformation identification on online social media platforms, providing opportunities for feature extraction, pattern recognition, and enhanced automated detection of deceptive content.

3.3 Hybrid model for detecting misinformation
In the field of misinformation detection (MID), a diverse array of techniques has been employed to identify various types of misinformation, including fake news, rumors, spam, trolling, false information, and disinformation. While individual deep learning (DL) models have been used extensively in MID, the increasing need for enhanced performance has led to the exploration of hybrid DL models. This section reviews related works in MID based on deep hybrid models, which incorporate elements such as Convolutional Recurrent Neural Networks (CRNN), Convolutional Restricted Boltzmann Machines (CRBM), Ensemble-Based Fusion (EBF), and Long Short-Term Memory (LSTM).

Convolutional Recurrent Neural Network (CRNN): Combining Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) in hybrid models is an increasingly popular approach. These models are tailored to process structured sequences with spatio-temporal properties, making them suitable for MID tasks. CRNNs, which combine CNN for visual feature extraction and RNN for sequential learning, have been effectively applied to fake news, rumor, false information, and spam detection. For instance, they enable the identification of rumors by utilizing hierarchical recurrent convolutional neural networks (RCNN) for contextual information and bidirectional GRU networks for temporal information.

Convolutional Restricted Boltzmann Machine (CRBM): A variation of the Restricted Boltzmann Machine (RBM), CRBM, is a two-layer model with organized visible and hidden factors as matrices. It effectively processes images and extracts features. CRBMs are employed in learning features specific to object classes, where associations are nearby, and weights are shared to construct multilayer progressive networks. These networks are crucial for image processing and feature extraction.

Ensemble-Based Fusion (EBF): EBF models consider additional profile information, such as speaker profiles, to enhance detection accuracy. They leverage crowd signals to detect fake news by analyzing user flagging behaviors. These models also detect deceptive words used by fake online users to cause harm offline. EBF-based methods make use of deep learning techniques, such as Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM), to counteract the increasing prevalence of misinformation.

Long Short-Term Memory (LSTM) Density Mixture Model: LSTM-based hybrid models have gained significant attention for MID tasks. These models effectively handle long sentences and employ attention mechanisms to weigh the importance of different words in context. They incorporate speaker profiles into attention-based LSTM models for fake news detection. LSTM architectures are used to exploit both content and metadata for bot detection, which is essential in addressing issues like political manipulation and stock market disruption. Additionally, hybrid DL models are used for the automatic identification of inappropriate language, contributing to improved conversation quality in various contexts.

4 Discussion with open issues and future research
In recent years, researchers have leveraged the advancements and accessibility of deep learning (DL) techniques to address issues related to misinformation, such as fake news, rumors, spam, and more, in online social networks. These technologies have provided powerful tools to process and analyze vast amounts of information. DL is becoming increasingly prominent in various domains, including business intelligence, predictive analytics, and learning management systems.

Despite the growth in DL applications, there remain several challenges and areas for improvement in misinformation detection (MID) research. The lack of large-scale publicly available datasets makes it challenging to compare different studies directly. To address these issues and advance the field of MID, researchers have explored various DL methods inspired by prior works. However, there are still critical issues to be addressed:


 * Semantics Understanding: Misinformation often involves semantics that are designed to deceive users. Achieving a deep semantic understanding is a significant challenge for machines. It is essential for MID to differentiate between various types of misinformation accurately.


 * Multimodal Data for Misinformation: Misinformation takes various forms, including text, images, and videos. Extracting features from each modality and understanding their interplay is a complex task. Comprehensive and large-scale datasets that encompass these various modalities are needed for improved MID.


 * Content Validation: Fact-checking and validating information quality is vital in combating misinformation. Developing high-quality fact-checkers and tools for crowdsourcing content validation can be valuable in identifying incorrect information.


 * Spreader Identification: Identifying influential users and spreaders of misinformation on social networks is crucial. Existing techniques do not always provide an accurate assessment of the nodal spreading capability or differentiate between the influence of various nodes.


 * Misinformation Identification: Most existing methods tend to focus on alerting users to misinformation without explaining why the information is false. Identifying misinformation even when users are not directly related is challenging.


 * Anomalous and Normal User Identification: Distinguishing between honest and dishonest users on online social media is a pressing issue. Detecting dishonest users who exploit the platform for their benefit can help maintain trust in online communities.


 * Bridging Echo Chamber: Bridging social media echo chambers that contribute to the spread of misinformation is essential. Data-driven models that reduce polarization and promote the exchange of opposing views need further research.


 * Mining Disinformation: The complexity and diversity of disinformation make it challenging to mine and detect effectively. Innovative DL strategies, such as reinforcement learning, can help address this issue.


 * Misinformation Dynamics: The spread of misinformation depends on various factors, including content, user behavior, and network structure. Dynamic models that capture the changing behavior of users can help combat the spread of fake news and misinformation.


 * Training on Large Data in Less Time: DL is known for its effectiveness with large datasets, but it often requires significant training time. Research into more efficient methods for training large data quickly can be beneficial.


 * Dynamic Data Analysis: Current studies primarily focus on static data analysis. The ability to analyze dynamic data, which is more representative of real-life situations, is an important direction for future research.


 * Explainable MID: Expanding the scope of MID to include explanations for why information is false can provide deeper insights and enhance the effectiveness of misinformation detection.


 * Integration of Reinforcement Learning: Combining reinforcement learning with DL techniques for misinformation detection can yield improved results, as reinforcement learning enables agents to learn from their interactions with the environment.


 * Addressing Overfitting: Overfitting is a common challenge in DL, impacting model performance in real-world scenarios. Addressing this issue is critical to ensure the reliability of MID models.

5 Conclusion
In this comprehensive survey of Misinformation Detection (MID) in social networks, we have extensively examined the various facets of false information, rumors, spam, fake news, and disinformation, elucidating their deceptive impact on online social platforms. Through our analysis, we have established that Deep Learning (DL) stands out as a potent and effective technique for addressing MID challenges. It not only excels in the identification of false facts but also performs at a level akin to human capabilities, making it a leading approach for analyzing social network data. Despite its strengths, several challenges persist, such as data volume, quality, interpretability, and privacy, which necessitate further exploration in future research. Our survey serves as a valuable resource for researchers and highlights the potential for incorporating user mental health conditions into the analysis of historical data, promising a more nuanced understanding of their online behaviors and misinformation-spreading tendencies, further advancing the field of MID.