key: cord-0666646-qm1298bs
authors: Meel, Priyanka; Vishwakarma, Dinesh Kumar
title: Fake News Detection using Semi-Supervised Graph Convolutional Network
date: 2021-09-28
journal: nan
DOI: nan
sha: a355abfab37a8daba7157434123ce8f17e175ee1
doc_id: 666646
cord_uid: qm1298bs

Social media becomes the central way for people to obtain and utilise news, due to its rapidness and inexpensive value of data distribution. Though, such features of social media platforms also present it a root cause of fake news distribution, causing adverse consequences on both people and culture. Hence, detecting fake news has become a significant research interest for bringing feasible real time solutions to the problem. Most current techniques of fake news disclosure are supervised, that need large cost in terms of time and effort to make a certainly interpreted dataset. The proposed framework concentrates on the text-based detection of fake news items while considering that only limited number of labels are available. Graphs are functioned extensively under several purposes of real-world problems on the strength of their property to structure things easily. Deep neural networks are used to generate great results within tasks that utilizes graph classification. The Graph Convolution Network works as a deep learning paradigm which works on graphs. Our proposed framework deals with limited amount of labelled data; we go for a semi-supervised learning method. We come up with a semi-supervised fake news detection technique based on GCN (Graph Convolutional Networks). The recommended architecture comprises of three basic components: collecting word embeddings from the news articles in datasets utilising GloVe, building similarity graph using Word Movers Distance (WMD) and finally applying Graph Convolution Network (GCN) for binary classification of news articles in semi-supervised paradigm. The implemented technique is validated on three different datasets by varying the volume of labelled data achieving 95.27 % highest accuracy on Real or Fake dataset. Comparison with other contemporary techniques also reinforced the supremacy of the proposed framework.

The constant extension of social media presented multiple parallel channels for users with added versatile facilities to obtain news than a few decades back. In accordance with the Pew Research Center [1] , nearly 2/3 of U.S. grown-ups made headlines of social media. Since people proceed to profit of the service and simple convenience from social media, they additionally reveal themselves to some captivating and incorrect news published on social media platforms particularly fake knowledge. These fake news stories are purposely composed to carry incorrect information as a kind of idea so as business or governmental manipulation.

The widespread of fake headlines could induce severe consequences on people and culture.

Thus, identifying and alleviating fake news becomes a critical obstacle in current social media themes. The term "fake news" can be described in varied configurations that slightly differs from each other but most of the times can be used interchangeably as: misinformation, disinformation, hoax, propaganda, clickbait, misconception, fabrication, deceit etc. [2] [3] [4] [5] as pictured in word cloud representation in Figure 1 . Graphs normally occur in various real-world problems employing cultural and social issues. Information that is accessible to us can be transformed in a structured way by means of connections between various objects to give promising insights regarding data [5] . One crucial issue in Artificial Intelligence problems is to draw meaningful relation between one part of data to other parts of data. Currently the best way to describe/represent these relations and connections between data is by employing graphs data structure that exist universally in modern technology solutions by providing us with a plan to mathematically describe complicated data as well as associated parameters.

Artificial Intelligence problem solving techniques can be broadly categorized into three classes: supervised, unsupervised and semi-supervised. Supervised technologies are well established to provide benchmark solutions in healthcare, data authentication, web services, security and many more fields. The realm of semi-supervised and unsupervised paradigm still needs to be judiciously explored. The major pitfall of supervised training methods are annotation inconsistencies, cost and manpower required due to massive data size. In order to create a good balance between labelled and unlabelled data we tried to use both at their best level of performance. Labelled data steers the training process towards right direction and unlabelled data is used to enhance model generalization as well as performance. Therefore, a good balance of both is highly required that we categorically attempted to incorporate in our suggested framework. Kusner et al [6] established Word Mover's Distance in 2015 as a distinct case of Earth Mover's Distance [7] and also proved that WMD gives unprecedented performance when coupled with K-nearest neighbour in classification tasks. This measure is suitable for calculating the distance between two text documents based on alignment between the words that are semantically close but syntactically different. It basically measures the dissimilarity between two text documents as the minimum distance the word vectors of one document need to travel to reach the word vectors of another document. GCN and its variants are being well tested to provide outstanding performance in variety of real-life applications incorporating both textual and visual data such as automatic seizure detection using temporal GCN [8] , spatiotemporal GCN for picking up and returning demand of public bike sharing by exploring information from multi-view data [9] , Autism Spectrum Disorder and Alzheimer's Disease prediction with imaging and non-imaging information [10] , semi-supervised hyperspectral image (HIS) classification framework using spectral-spatial graph convolution network (S 2 GCN) [11] , multi-stream attention-enhanced adaptive graph convolutional neural network (MS-AAGCN) for skeleton-based action recognition [12] etc.

The major problem we face in classification techniques is less no. of available labels for data. Apart from this cost, manpower and labelling inconsistencies are major concerns for annotating huge volumes of news articles. Thus, we make use of semi-supervised learning method. Graphs are capable of representing structural and contextual dependencies between its nodes, each representing an object and subsequently can be utilized in cases where the number of labels is limited [13] . This encourages us to take benefits of Graph Convolution Network for classification. Given one node, for all the classes can virtually start training procedure. Such a model, can very quickly lead to the creation of a 2d vector space with word embeddings involved which could be visualized easily. As a result, it can be seen that with just 3 layers in the architecture along with one label per actual class and absolutely no features in place, a linear classification is made. This produces quite exceptional results with given constraints and conditions. The key contributions of proposed framework are as follows:

 The architecture presents a semi-supervised fake news detection technique based on GCN (Graph Convolutional Networks) trained with limited amount of labelled data.

 Elaboratively elucidate three building blocks of the framework: collecting word embeddings from the news articles in datasets utilising GloVe, constructing similarity graph using Word Mover's Distance (WMD) and finally applying Graph Convolution Network (GCN) for binary classification of news articles in semi-supervised paradigm.

 Graph Convolution Network can harness the best advantage of convolutions as well as data structuring capabilities of graphs to draw meaningful insights out of complicated data and associated parameters.

 The implemented technique is validated on three different datasets by varying the volume of labelled data. 

The subject of fake news becomes an emerging issue within modern social media soundings. Current veracity analysis methods can be broadly classified into a couple of sections [13] , [14] : utilising news data for handling social information/context, extracting linguistic, syntactic, lexical features or visual characteristics [15] , [16] from news articles to obtain particular features as well as interesting patterns that usually happen in fraudulent articles.

Visual features are adopted to recognise fake pictures that are deliberately produced to exert realistic effects in fake news. Context-based approaches of veracity analysis includes characteristics of user-data on social sites and social systems [17] . User social site's profile is utilised to estimate that users' features and reliability. Characteristics derived from their social posts describe their social behaviour. Pathak et al. [18] presented elaborated discussions on characteristics of publicly available datasets for veracity analysis techniques and performance comparison with contemporary machine learning, deep learning, hybrid approaches of supervised and unsupervised domain. Contemporary survey of semi-supervised and unsupervised methods of web content credibility analysis is detailed by Saini et al [19] . We studied and categorized the available literature in the domain of veracity analysis into three different classes: supervised, semi-supervised and unsupervised methods. Apart from this also highlighted the unprecedented performance of different variants of GCN in a variety of reallife applications including healthcare, recommendation systems, traffic management, ecommerce, retail, marketing etc. that serves as a motivation to utilize it further for our designed fake news detection framework.

The supervised methods principally concentrate on selecting useful characteristics and utilise them to develop supervised training structures. A neural network-based Graph-aware Co-Attention Networks (GCAN) [20] framework considers realistic scenario of social network to predict whether a source tweet is fake or real at the same time producing the reasonable explanations as well as highlighting evidences on suspicious tweeters. The network requires the input in the form of short text tweets and the corresponding sequence of retweet users without text comments. GCAN significantly outperforms contemporary techniques by 16% in accuracy on average on Twitter 15 and Twitter 16 datasets. To model the sematic representations of fake news by jointly modelling the textual representations, knowledge concepts and visual information into a unified framework Wang et al. [21] proposed a novel Knowledge-driven Multimodal Graph convolutional Network (KMGCN). After knowledge distillation process a well-designed graph structure is used to effectively represent textual, knowledge and visual information to apply GCN succeeded by pooling layer, fully connect layer and finally softmax function for classification. The framework is experimented and validated on WEIBO and PHEME datasets. A weakly supervised reinforcement learning framework (WeFEND) proposed by Wang et al. [22] comprises of three components:

annotator, reinforced selector and fake news detector. The annotator automatically assigns weak labels for news articles, the reinforced selector filters out high-quality samples using reinforcement learning and the fake news detector classifies fake articles based on content of news. Wu et al. [23] proposed a graph-based method of rumour detection in which a propagation graph is being constructed based on who replies to whom relationship on Twitter.

Two models GLO-PGNN (global embedding with propagation graph neural network) and ENS-PGNN (ensemble learning with propagation graph neural network) adopting diverse classification approaches for rumour detection task are designed. Both the methods employ attention mechanism for dynamically adjusting the weights to improve performance. Meel and Vishwakarma [24] proposed multimodal fake news detection method based on ensemble of hierarchical attention network, image captioning and forensics analysis. Raj and Meel [25] experimented with Resnet50 , VGG16, VGG19,InceptionV3, DenseNet, Xception, AlexNet , MobileNet deep arcitectures on T1-CNN, EMERGENT, MICC-F220 multimodal datasets to compare and contrast the capacities of different ConvNet architectures for textual as well as visual forgery detection.

A semi-supervised naïve architecture Factual News Graph (FNG) [26] based on graphical social context representation is proposed that demonstrates significant improvements in fake news detection task and robustness in the presence of limited training data. The FANG representation learning framework is optimized by three concurrent losses unsupervised proximity loss, self-supervised stance loss and supervised fake news detection loss. Guacho et al. [27] recommended semi-supervised misinformation detection in graphs via embedding the input articles in multi-dimensional tensors which are further decomposed to capture spatial and contextual information by creating article-by-article-graph. A temporal ensembling based semi-supervised fake news detection framework using self-ensembling of the combined prediction of all previous epochs to serve as a proxy label for the current epoch is proposed by Meel and Vishwakarma [2] . Online opinion reviews have an economic impact on goods and services. Opportunistic firms often manipulate online reviews to make undue profits by unethically influencing people's choices. Rout et al. [28] proposed a semi-supervised method of classifying deceptive and fake opinion reviews and validated the framework on hotel reviews. A novel semi-supervised framework for multimodal fake news detection is designed by Mansouri et al. [29] using Linear Discrimination Analysis (LDA) and Convolutional Neural Network (CNN). Li et al [30] proposed a self-learning semi-supervised deep learning network that adds a confidence network layer to made it possible to automatically return and add correct results for improving the accuracy of the neural network.

Annotated samples cannot characterize the authenticity of the news on recently developed event as they become obsolete very quickly due to the dynamic nature of news. To obtain highquality up-to-the-minute labelled sample is the foremost challenge in supervised and semisupervised learning methodologies. The motive behind development of unsupervised technologies is to completely eliminate the need of annotation and real time solution to the problem of information trustworthiness. A novel unsupervised method of detecting fake news circulating on WhatsApp is implemented by Gaglani et al. [31] leveraging Transfer Learning techniques. The framework utilizes semantic similarity between claims circulated on WhatsApp and associated articles scrapped from web to classify the claims as true or false.

Hosseinimotlagh and Papalexakis [32] recommended an unsupervised ensemble method that consolidates results from different tensor decompositions into coherent, comprehensible and high precision groups of articles that belong to different categories of false news. Unsupervised microblog rumour detection framework is proposed [33] based on recurrent neural network and autoencoders by analysing the temporal dynamics and crowd wisdom. Yang et al. [34] recommended a generative approach of unsupervised fake news classification on social media by exploiting user's engagement, opinion and their credibility factors. A three phased graphbased approach utilizing biclique identification, graph-based feature vector learning and label spreading for fake news detection in the absence of annotated data is successfully designed by Gangireddy et al. [35] .

Singh et al. [36] . Graph Convolution Network is being explored by Yao et. al [37] in conjunction with LSTM network to explore between objects in an image based on their spatial and semantic connections for generating image captions. Jeong et al. [38] proposed a deep learning based context aware citation recommendation framework encompasses a document encoder and content encoder using GCN and BERT .Other remarkable contributions of GCN for designing solutions to challenging problems can be highlighted as: semi-supervised hyperspectral image (HIS) classification framework using spectral-spatial graph convolution network (S 2 GCN) [11] , Infusing knowledge into the textual entailment tasks [39] , multi-label image classification framework using GCN [40] , robot navigation in crowds using GCN with attention trained by human gaze data that accurately predicts human attention to different agents in the crowd [41] etc.

The inspirations for developing semi-supervised fake news detection framework and prime research targets can be regarded as follows:

To design a semi-supervised Fake News Detection framework that can learn from the semantics of labelled data and intrinsic patterns of unlabelled data to optimize in terms of time, cost, labour and annotation inconsistencies.

To address the highly circulated forgery format of a news story coupled with headline text along with partially labelled training samples.

To take advantage of self-learning cognition of designed framework which facilitates the architecture to learn from up-to-the minute latest forgery formats fed into the process in the form of unlabelled articles.

Kusner et al [6] established the fact that Word Mover's Distance gives unprecedented performance when coupled with K-nearest neighbour in classification tasks.

 Graph Convolution Network can harness the best advantage of convolutions as well as data structuring capabilities of graphs are capable of representing structural and contextual dependencies between its nodes (documents) to draw meaningful insights out of complicated data and associated parameters.

The framework is aimed to avoid cumbersome feature engineering of supervised learning methods, crowdsourcing or human expert annotation inconsistencies, reduces cost and manpower requirements.

The semi-supervised framework of Graph Convolutional Network combines best features of convolutions and data modelling capabilities of graphs. The proposed architecture incorporates a three-step approach to classify the articles as fake or real. First step is to transform the given textual data from a dataset in the form of vectors. These vectors represent the linguistic features of the text and can be utilized to represent an article in Euclidean Space.

Global Vector (GloVe) embedding is used to transform the articles into vectors. Each article is interpreted as the mean of vectors of the words it contains. The result of this embedding is used to construct a similarity graph between articles in the dataset. Each node represents an article and most similar nodes are connected by an edge in the graph. To calculate the similarity, i.e. distance between two articles, Word Mover's Distance is used. It uses semantically meaningful relations between words to find the similarity between two articles. Before converting text into vectors basic pre-processing steps such as: dropping unused columns, removing null and missing data, removing stope words, tokenizing the articles, lemmatizing each word, converting labels name etc. are done for all the datasets. The subsequent procedure involves word embedding and similarity graph construction. This results in creation of graphs containing three types of nodes labelled as real, fake and unlabelled which can be inputs to the graph convolutional neural networks for finally predicting the output. The process is pictorially emphasized in figure 2 and also comprehensively detailed in algorithm1. 

Article embedding in Euclidean Space means transforming text to multidimensional vectors by using word embedding. Word embedding is the process of constructing mathematical equivalents, i.e., representation in the mathematical form of each word from some corpus. It is implied that words with similar or same sense will have close representations.

In Natural Language Processing, word embedding has proved to be one of the concepts that have huge applications, as it lays the foundation for solving many real-world tough problems.

The process involves constructing vectors for each word in any number of dimensions such that it can be visualized as a vector in a vector space in those number of dimensions. This conversion of words to vectors uses neural network methodology to find the values of vectors.

The vector can have any number of dimensions ranging from a single digit to a few hundred.

As words with similar or same sense will have close representations, it can be implied that the meaning of words has been learnt by a model, which in turn can be useful for solving challenging problems. Our method will utilize the final word vectors created by the embedding of words for each article by representing it in an euclidean space. Further this will also be converted into low dimensions using a dimensionality reduction technique and contextual associations in news articles will be obtained by using a similarity graph representation.

To create a vector which maps to each news article, our proposed semi-supervised method uses a pre-trained GloVe model of 300 dimensional embeddings and for all the words appearing in that article we compute a mean vector which is finally mapped to that article. 

The classification step requires a graph input, which is done by constructing a similarity graph representing closeness in between different nodes, i.e., each news article. Such a graph can be represented as a k-nearest neighbour graph. In particular, for every news article, we look forward towards the k-nearest neighbours in the embedding space. The k neighbours can be determined by calculating the Word Mover's distance between one article to all the other news articles.

A k-NN (k-nearest neighbour) graph is one in which node p and node q have an edge between them if node p is one of the top k nearest of all its neighbours or vice-versa. Every knearest-neighbours of a location in the n-dimensional space will be determined to utilise a "closeness" relationship wherever closeness is usually described within words of a distance measure as Word Mover's Distance. Thus, with given articles in a vector space, a k-NN graph of points can be created by calculating the remoteness between each pair of points and connecting each point with the k most proximal ones.

The Word Mover's Distance (WMD) proposed by Kusner et al. [6] in 2015 is a distance between documents that takes benefit of semantic relations among words that are apprehended by their embeddings. It is a special case of Earth Mover's Distance [7] and also provides extraordinary performance when coupled with K-nearest neighbour in classification tasks. It basically measures the dissimilarity between two text documents as the minimum distance the word vectors of one document need to travel to reach the word vectors of another document.

This distance proved to be quite effective, obtaining state-of-art error rates for classification tasks.

WMD metric is a distance function between text documents that leads to unparalleled low k-nearest neighbour document classification error rate. Precisely representing the remoteness between two documents has far reaching utilities in document retrieval, multi-lingual document matching, text clustering etc. To calculate a distance between two text documents the basic unit is "travel cost" between two words. Let "A" and "B" be the representations of two text documents with |A| and |B| are the number of unique words in both the documents respectively.

Each word in document A is allowed to be converted into any word in document B. ≥ 0 denotes the amount of distance word in A has to travel to convert into word in B. T is a sparse transportation flow matrix. To transform A completely into B we have to confirm that the complete outgoing flow from i th word equals and incoming flow to j th word must match represented mathematically as equation 1.

The WMD distance between two documents is the minimum weighted cumulative cost required to move all words from document A to document B represented as

Formally the WMD between two documents is defined as the value of the optimal solution of the following transportation problem which is a special case of Earth mover's distance.

(3)

Word Mover's Distance (WMD) is derived upon current events in embeddings of the words which determine the semantically significant description of those words of local co-occurrences within those sentences among new articles. Salient features of Word Mover's Distance can be characterized as:

 It is easy to learn, practice and free from the effect of any hyperparameters.

 It can be easily described as the distance among two text contents that can be split and described as the distances which are sparse within different words.

 It commonly includes some information represented as the word2vec or Glove and heads to huge operations accuracy.

Graph Convolutional Network is a resourceful development of convolutional neural networks which functions directly on graphs. The model scales linearly in the number of graph edges by using a competent layer-wise propagation rule that is based on first-order approximation of spectral convolutions on graphs. GCN model is capable of learning hidden layer representations that encodes both node feature and graph structure to an extent useful for semi-supervised classification. result when compared to a traditional learning algorithm. Figure 4 illustrates the basic design pipeline for a GCN model. Neural networks apply non-linear activation functions to represent the non-linear features in latent dimensions. Forward pass equation in neural network is represented as in equation (7) where σ represents activation function, ( +1) and ( ) represents feature representation at l th and (l+1) th layer, ( ) weight and ( ) bias at layer l.

( +1) = ( ( ) ( ) + ( ) )

Forward pass equation in Graph Convolutional Network can be represented as: A is adjacency matrix representing the connections between the nodes in graph structured real world data. Adjacency matrix A enables the model to learn the feature representations based on nodes connectivity. Bias b is omitted to make the model simpler. Equation (8) is the first-order approximation of spectral graph convolution propagating the information along the neighbouring nodes within the graph.

The aim of Graph Convolutional Network framework is to learn a function of features on a graph which takes as input a feature matrix X of dimension NxD where N is the number of nodes and D is number of input features. A is the adjacency matrix for representing the overall graph structure. A and X are the input to GCN architecture producing the node level output Z (an NxF feature matrix, where F is the number of output features per node). A nonlinear function representing every neural network layer in GCN can be written as in equation (9) and equation (10) highlights layer wise propagation across the network.

With (0) =X is initial input, ( ) =Z is node level output at last layer (or z for graph level output) , L is the number of layers , ( ) is a weight matrix for the l th neural network layer and σ is a non-linear activation function. Instead of using matrix A for semi-supervised classification in practical scenario Kipf and Welling [43] proposed to use symmetric normalization D -1/2 AD -1/2 , so equation (9) can be rewritten as:

With ̂ = A+I, I is identity matrix and ̂ is the diagonal node degree matrix of A. In the process of semi-supervised classification with GCN the network initially starts training on the labelled nodes, subsequently propagating the information to unlabelled nodes by updating weight matrices that are shared across all nodes. This process can be summarized in the following steps:

 Accomplish forward propagation through GCN.

 Put on the sigmoid function row-wise on the last layer in the GCN.

 Calculate the cross-entropy loss on known node labels.

 Backpropagate the loss and update the weight matrices 'W' in each layer.

In this sub-section, we extensively detail the experimentation settings, datasets, parameter 

Three different fully labelled datasets Fake News Data, Real or Fake and Fake News Detection introduced on Kaggle platform are used for experimentation and validation of our work. The headline, body and label part of each one of the datasets are being utilized for model training and testing purpose. The details of the datasets are listed in following Table 1 : 

Performance of the proposed framework on three different datasets is being evaluated and compared for accuracy, precision, recall and F1 score. The experiments for each dataset have been repeated for two different values of K (K=3, K=5) and varied proportions of labelled training data ranging from 20% to 50%. 

To compare the efficacy of our designed architecture with contemporary methods, we calculated the performance of five different baseline methods in terms of accuracy, precision, recall and F1 score with the dataset split as 8:1:1 for training, validation and testing. State-ofthe-art juxtaposition on each of the Fake News Data, Real or Fake and Fake News detection datasets are outlined in Table 5, Table 6 and Table 7 , correspondingly. The approaches used as baselines for state-of-the-art comparison are as follows:

 Bali et al. [47] proposed Support Vector Classifier (SVC), Random Forest(RF), Naïve The above discussion concludes that our proposed model for veracity analysis of web information is reasonably promising. The precision over all the three datasets Fake News Data, Real or Fake and Fake news detection is a decent development over the parallel methods. Graph Convolutional Network and Word Mover's Distance for calculating distance are the prominent technologies that have facilitated refining the preciseness of our results. Finally, experimenting with different values of K and varied proportions of labelled data helps us achieve 95.27%

highest fake news detection accuracy.

Fake news is an increasing concern for the modern society. One of the most difficult part in providing a solution to this is labelling massive volumes of data to train supervised artificial intelligence models. To counter this challenge, we tried to design a semi-supervised text fake news detection framework based on Graph Convolutional Networks. Embedding the text articles in Euclidean space, similarity graph constructing using Word Mover's Distance and Graph Classification are the three landmark components of the designed architecture.

Extensive experimental analysis has been done by repeating the training and testing of the model for different proportions of labelled and unlabelled data for each dataset. The promising results compared in terms of accuracy, precision, recall, F1 score and ROC Curves very well advocates the worthiness of the method for veracity analysis. The proposed framework exhibits encouraging results by outdoing numerous state-of-the-art methods on multiple standard datasets.

In spite of all the efforts done by researchers, governments and corporates to put an end to this malice of fake news, a lot more is still to be done. Temporal component is extremely important in current real-life decision-making process. Correct information received too late is also irrelevant, so real time information authentication systems are highly in demand. Platforms such as Facebook, WhatsApp, LinkedIn, Gmail etc. are implementing solutions for user privacy as well as data security and trustworthiness. The research could be further extended to incorporate multimedia data as visual data constitute an integral part of digital online information. Advanced deep learning technologies of transformers and transfer learning can also be amalgamated with graph data structures to make the process more efficient. Stand-alone modules that can be added as browser and application plugins could also be the subsequent research tracks. The classification performance of fake news detection systems can also be improved by using attention mechanism instead of complex convolutions as well as development of unsupervised frameworks on large scale unlabelled data further enhances the worthiness of the procedure.

Pew Research Center

A temporal ensembling based semi-supervised ConvNet for the detection of fake news articles

A review on rumour prediction and veracity assessment in online social network

A unified approach for detection of Clickbait videos on YouTube using cognitive evidences

Recent State-of-the-art of Fake News Detection: A Review

Semi-Supervised Learning and Graph Neural Networks for Fake News Detection

From word embeddings to document distances

The earth mover's distance as a metric for image retrieval

Temporal graph convolutional networks for automatic seizure detection

Demand prediction for a public bike sharing program based on spatio-temporal graph convolutional networks

Disease prediction using graph convolutional networks: application to autism spectrum disorder and Alzheimer's disease

Spectral-spatial graph convolutional networks for semisupervised hyperspectral image classification

Skeleton-based action recognition with multi-stream adaptive graph convolutional networks

Classification of Propagation Path and Tweets for Rumor Detection using Graphical Convolutional Networks and Transformer based Encodings

A survey on fake news and rumour detection techniques

Combating fake news: A survey on identification and mitigation techniques

Fake news, rumor, information pollution in social media and web: A contemporary survey of state-of-the-arts, challenges and opportunities

Rumor Propagation: A State-of-the-art Survey of Current Challenges and Opportunities

Polarization and Fake News: EarlyWarning of Potential Misinformation Targets

Analysis of Techniques for Rumor Detection in Social Media

Multimodal, Semi-supervised and Unsupervised web content credibility analysis Frameworks

GCAN: Graph-aware Co-Attention Networks for Explainable Fake News Detection on Social Media

Fake News Detection via Knowledge-driven Multimodal Graph Convolutional Networks

Weak supervision for fake news detection via reinforcement learning

Rumor detection based on propagation graph neural network with attention mechanism

HAN, image captioning, and forensics ensemble multimodalfake news detection

ConvNet frameworks for multi-modal fake news detection

FANG: Leveraging social context for fake news detection using graph representation

Semi-supervised Content-based Detection of Misinformation via Tensor Embeddings

Revisiting semi-supervised learning for online deceptive review detection

A Semi-supervised Learning Method for Fake News Detection in Social Media

A novel self-learning semi-supervised deep learning network to detect fake news on social media

Unsupervised WhatsApp Fake News Detection using Semantic Search

Unsupervised content-based identification of fake news articles with tensor decomposition ensembles

Unsupervised rumor detection based on users' behaviors using neural networks

Unsupervised Fake News Detection on Social Media: A generative approach

Unsupervised Fake News Detection: A Graph-based Approach

State-of-the-Art Applications of Graph Convolutional Neural Networks

Exploring visual relationship for image captioning

A context-aware citation recommendation model with BERT and graph convolutional networks

Infusing knowledge into the textual entailment task using graph convolutional networks

Multi-label image recognition with graph convolutional networks

Robot navigation in crowds by graph convolutional networks with attention learned from human gaze

GloVe: Global Vectors forWord Representation

Semi-supervised classification with graph convolutional networks

Fake News Data

real_or_fake

Fake News Detection

Comparative Performance of Machine Learning Algorithms for Fake News Detection

Fake News Detection using Machine Learning and Natural Language Processing

Learning Hierarchical Discourse-level Structure for Fake News Detection

Detection and veracity analysis of fake news via scrapping and authenticating the web search