Impact of Transformer-Based Models and User Clustering in Early Fake
News Detection in Social Media
Sakshi Kalra
1
, Yashvardhan Sharma
1
, Mehul Agrawal
1
, Sai Ratna Kashyap Mantri
1
and Gajendra Singh Chauhan
2
1
Department of CSIS, BITS Pilani, Pilani, 333031, Rajasthan, India
2
Department of HSS, BITS Pilani, Pilani, 333031, Rajasthan, India
Keywords:
Early Fake News Detection, Neural Networks, Transformers, Attention Mechanism, User Clustering, Fuzzy
C-Means Clustering, K-Means Clustering.
Abstract:
People are now consuming news on social media platforms rather than through traditional sources as a result of
easy access to the internet. This has allowed for the recent rise in the online dissemination of false information.
The spread of false information seriously damages people’s reputations and the public’s trust in them. The
research community has recently given fake news identification a great deal of attention, and prior studies have
mainly concentrated on finding hints in news content or diffusion graphs. The older models, on the other hand,
didn’t have the key features needed to spot fake news quickly. We focus on finding fake news by using features
that are available when it is just starting to spread. The current work suggests a new framework made up of
content-based features taken from news articles and social-context features taken from user characteristics
and responses at the sentence level. In addition, we extend our approach to Transformer-based models and
leverage user clustering to demonstrate a considerable performance gain over the original model.
1 INTRODUCTION
Dealing with fake news has been part of our daily life
in recent years. The spread of misinformation can
heavily hamper a person’s personal fame and public
trust. Social media sites like Twitter and Facebook
make it easier for people from all over the world to
share information in real time. It has become the main
way that people connect and share information on-
line because it is easy to use, doesn’t cost much, and
moves quickly. With the popularity of social media
growing so quickly, the internet has become a place
where fake reviews, fake political statements, fake
news, etc. are all over the place. For example, ar-
ticles stating “COVID-19 vaccination causes autism
and infertility among recipients
1
” can essentially im-
pact the public trust and may prompt a drop in im-
munization drives. As per the research, fake news
spreads much faster and deeper than factual news
(Vosoughi et al., 2018). This has drawn significant at-
tention among the industry leaders and research com-
munity as well. Even though the main motto of so-
cial media is to provide better communication, many
users have started to confuse news from such plat-
1
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9359307/
forms with main stream media. Traces of fake news
started in 1439 itself (Klyuev, 2018) but the ease and
scale at which it was disseminated changed drastically
over time. In the modern day of the internet, fake
news gained its importance only during the 2016 US
presidential elections (Kshetri and Voas, 2017). There
is no concrete definition of fake news. Based on ex-
isting literature, fake news can be loosely defined as
“false stories that appear to be news and spread on the
internet or other media, usually created to influence
political views or as a joke, and depicting deliberate
intentions”. This depicts why fake news spread so
rapidly, because most of the online news publishers
have poor credentials and deny to identify themselves,
which creates room to spread misinformation.
Early research on spotting fake news was mostly
about finding better ways to spot fake news. This
included, but wasn’t limited to, understanding con-
text, how news spreads, writing styles, syntactic anal-
ysis, etc. Getting such useful features is often hard
and takes a lot of time, but users are smart enough to
find ways around this. Recent research has focused
on solving the above-mentioned problems. To learn
how to represent a diffusion graph, for example, a lot
of attention is paid to matrix factorization, graph neu-
ral networks, recurrent neural networks, and convo-
Kalra, S., Sharma, Y., Agrawal, M., Mantri, S. and Chauhan, G.
Impact of Transformer-Based Models and User Clustering in Early Fake News Detection in Social Media.
DOI: 10.5220/0011684000003411
In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 889-896
ISBN: 978-989-758-626-2; ISSN: 2184-4313
Copyright
c
 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
889

lutional neural networks. These methods only work
well when they have enough information about how
fake news is spreading. They are not good enough to
find fake news in its early stages. As a human, when
we are given a piece of information, we first use our
intuition to judge its factual correctness. At times, we
might also look for a reliable source to verify the in-
formation. This scenario motivates the importance of
publisher and user credibility in detecting fake news
in much earlier phases.
2 RELATED WORK
The detection of fake news on social media has drawn
a lot of attention in recent years. One of the main
goals of the studies that have already been conducted
is to create machine and deep learning-based clas-
sifiers that can automatically tell if a news article
spreading on social media is fake based on a number
of news features. Early research focused on finding
linguistic clues in news articles that could be used to
spot fake news. This section gives an overview of the
research that has been done on automatic extraction of
the features for spotting fake news and closely related
topics like spotting rumors or misinformation.
2.1 Analysis Related to News Content
Based Features
Many researchers use the simple method of just look-
ing at the news content to spot fake news. They
read the news article headlines, bodies of text, and
in some cases, related images and videos (Jin et al.,
2016). Some, such as (Gupta et al., 2014) counted
the number of swear words and words that contained
pronouns in order to create features to distinguish
fake news from real news. (Castillo et al., 2011)
adopted a list of content-based features, including
emoticons, pronouns, sentiment of words, and punc-
tuation marks, used to determine the veracity of news.
Based on writing styles, (Afroz et al., 2012) found
online fraud, deception, and hoaxes. They have used
things like assertive verbs, factive verbs, and implica-
tives to figure out how likely web claims are to be true.
These stylistic linguistic features can be easily manip-
ulated and do not convey semantic meaning. These
methods therefore have a lower likelihood of being
successful in practical applications. Content-based
detection methods (Sun et al., 2013) often have trou-
ble finding fake news because it comes in many differ-
ent forms, in many different ways, and on many dif-
ferent platforms. Additionally, news content features
may be event-specific. As a result, features based on
content that perform well on one dataset of fake news
may not perform well on another.
2.2 Analysis Related to Social Context
Based Features
Social interactions related to a news article are in-
cluded in social context features. They might reveal
information about whether a news article is accurate.
Some research has already been done on the ways that
social context is used to classify news. The most com-
mon types of social context features are based on the
user, on the text, and on the structure. User profiles
on social media, which show what kind of people use
social media, can be used to get information about
user-based features. (Castillo et al., 2011) used a list
of fundamental user-based features supported by var-
ious social media platforms, such as the number of
followers, friend count, and age of registration, to as-
sess the accuracy of information posted by its source
user. (Yang et al., 2012) added a few user features to
Sina Weibo, a Chinese social media platform, in ad-
dition to the typical user characteristics, such as gen-
der and registration area, to find rumors. Using only
user-based features to decide if a news article is fake
has a big drawback: people who make fake news of-
ten mix it with real news to make it more likely that
people will believe it. So, even if the news article
isn’t true, just looking at how people use a resource
doesn’t give us a full picture. Information on the
users who shared or retweeted a news article, how-
ever, may give us more insight into the authenticity
of a news article. However, this type of feature is
ignored by many existing studies. Text-based social
context features can be accessed through the com-
ments and discussions of social media users that show
up under news articles. A number of temporal-based
features extracted from the time series of user com-
ments and time-stamped on user comments are pro-
posed to detect false news. (Ma et al., 2015) used
a time series of content and features based on social
context, such as the percentage of microblogs with
URLs and the percentage of verified users, to tell the
difference between rumors and other types of content.
But these ”aggregated level” parts need a lot of sta-
tistical considerations in order to spot fake news as
soon as it comes out. Many deep learning techniques,
like RNN, are used by (Ma et al., 2016) to extract
temporal-linguistic patterns from user comment se-
quences in order to identify rumors. At the beginning
of the news propagation process, user responses may
be very limited, which can have a significant nega-
tive impact on the performance of RNN models and
lead to them becoming overfit. This is one of the
ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods
890

main disadvantages of these methods. Social media
users can connect with one another through directed
or undirected links, such as friendship and following.
When a news story is shared through these links, a
propagation network can be formed. Existing stud-
ies have examined structural features extracted from
propagation networks as a different method of iden-
tifying fake news. For rumor detection, (Yang et al.,
2015) took advantage of the topology property of so-
cial networks. By comparing the diffusion patterns of
rumors and non-rumors, (Liu et al., 2017) were able to
identify rumors. The disadvantage of using structural
features is that it typically requires a lot of time to
observe a propagation network big enough to extract
useful distinguishable features, so these approaches
are not very effective in cases of early detection.
2.3 Research Objectives
Through this research work, we looked into the prob-
lems below and attempted to provide solutions using
our suggested model and techniques.
1. How to minimize the generalization problem of
content based models?
2. How to overcome the limited availability of struc-
ture based social context-features?
3. How to capture local and global variations in so-
cial context features?
4. How would an ensemble of content and social
context features improve the performance?
3 DATASET USED
The dataset used for this study is the FakeNewsNet
dataset described by (Shu et al., 2020). It combines
information from Twitter about the users who tweeted
the articles with a set of real and fake articles from
PolitiFact and GossipCop that were manually labeled
by humans.
4 PROPOSED METHODOLOGY
At the beginning of the spread, we have news content,
user profiles of people who share news on social me-
dia, and their tweet responses, which we get within
5 minutes of the news being posted. We can use all
of the available features in the early stages of propa-
gation by combining the limited social context-based
model with the content-based model.
4.1 Content Based Models
Two content-based models are used. One is based on
glove embeddings and convolution neural networks
to extract latent features, and the other is based on
transformers, which are state-of-the-art in many NLP
tasks.
4.1.1 Glove and CNN Based Feature Extraction
Firstly, we prepare a 3D input for a convolutional-
based model. The idea isn’t to get an embedding for
the whole article, but to break it up into sentences and
add the number of sentences as a third dimension to
the input. The main advantages of breaking down an
article into sentences are:
1. As each sentence in an article is represented by
a separate feature vector in the input tensor, we
can encode the positional information of the sen-
tences.
2. Other extra features at sentence level are also
combined using this methodology.
So, we transform the news headline and body into a 3-
dimensional tensor, where the headline and sentences
represent the first dimension, words represent the sec-
ond dimension, and glove word vectors represent the
third dimension. Then, we feed this input into a con-
volutional network to get a single feature vector for
the whole article. Figure 1 shows the 3D input ten-
sor. This 3D input tensor’s size must be fixed, so two
Figure 1: 3D Input Tensor.
thresholds are set: one to limit the number of sen-
tences in an article ((T
d
) and the other to limit the
number of words in a sentence ((T
s
). Any article hav-
ing sentences longer than (T
s
) is truncated, and lesser
ones are padded. Similarly, for sentences extending
the word limit, truncation is done; otherwise, they are
padded to the word limit. Based on how the model
was built and some statistical analysis of the dataset,
we chose (T
d
) = 100 (About 5% of sentences have
more than 100 sentences). For calculating (T
d
), we
obtained the mean no. of sentences in an article (µ)
and its standard deviation from the mean (σ) added
them to obtain the value of (T
d
). (T
d
) = (µ +σ). This
Impact of Transformer-Based Models and User Clustering in Early Fake News Detection in Social Media
891

Figure 2: Glove and CNN Based Content Feature Vector
Extraction.
prevented the construction of very large and sparse
tensors as it ignored the outlier sizes of articles in the
dataset. By doing similar statistical analysis, we ob-
tained (T
w
) = 46. The architecture of the convolution-
based network can be observed in Figure 2. In the
input layer, the whole news article is represented as
a 3D tensor. Then there are four horizontal convolu-
tional blocks (HCB), using which we extract one fea-
ture vector for each sentence, thereby obtaining a ma-
trix of size (100, 32) and then flattening and passing
it through a dense layer (64) to obtain a single feature
vector for the entire article. HCB is made up of two
convolution layers that come one after the other, fol-
lowed by a ReLU layer and then a max pooling layer.
This content feature vector is concatenated with the
social context-based feature vector to obtain the final
feature vector, which can be used to classify news as
fake or real.
4.1.2 Transformers Based Feature Extraction
Transformers-based models are the state-of-the-art in
various Natural Language Processing tasks. They
have a deeper understanding of the language and
have been pre-trained in both directions on large
datasets. We developed contextual embedding rep-
resentations for each sentence in an article and hence
obtained a 2D content feature matrix having a dimen-
sion of (number of sentences(100) * embedding vec-
tor dim(768)) as shown in Figure 3.
Later, this content feature matrix was passed
through a stack of dense layers and a few horizon-
tal convolution blocks to obtain a single feature vec-
tor for the entire article. Since it also extracts feature
representation for each sentence in the article, like
Figure 3: Transformer Based Content Feature Vector Ex-
traction.
SLCNN , it also encodes the positional information
of the sentences, hence including features at the sen-
tence level. The RoBERTa Base was used for finding
embeddings of the sentences since it yielded the best
accuracy and an F1 score as shown in Table 1.
Table 1: Comparative Analysis based on Various
Transformer-based Architectures.
Evaluation
Parameter
bert-base distilbert-
base
XLM
RoBERTa
RoBERTa-
base
Accuracy 78.7% 77.3% 75.5% 81.2%
F1-Score 76.3% 76.2% 73.5% 80.5%
4.2 User and Social Context Based
Model
User profiles of news spreaders on social media and
their tweet responses when they posted their tweet are
considered. For each article only K tweets are con-
sidered by the assumption that we will get those K
tweets in the first 10-15 minutes of the tweet being
posted. In that case we were able to get an average
of 10 tweets within 10-15 minutes of posting, hence
we used K=10. We have used this constraint of using
only limited number of tweets to ensure early detec-
tion of fake news.
4.2.1 Extracting Tweet Feature Matrix Using
Glove-Based Architecture
For each news article we consider 10 tweets and hence
10 tweet responses, each tweet response has an aver-
age of 15 words and each word will be represented
by glove vector, hence dimension of our input vec-
tor would be : (K * max no of words in a tweet re-
sponse * glove word vector size). We have used a
ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods
892

Figure 4: Tweet Feature Matrix Extraction using Glove-
based Architecture.
CNN based model to extract35 tweet feature matrix
from this 3D input tensor. It uses few horizontal con-
volutional blocks along with reshape layer in the end
to convert it into a 2D matrix of size (10 * 32) i.e.,
each tweet response being represented by a feature
vector of size 32. Figure 4 shows the tweet feature
matrix extraction using glove-based architecture.
4.2.2 Extracting Tweet Feature Matrix Using
Transformer-Based Architecture
We developed contextual embedding representations
for each tweet. Hence it generates a 2D input having
dimension: (no of tweets * embedding vector dim).
This matrix is further passed through a series of dense
layers to reduce the dimension of embedding vector
and obtain the tweet feature matrix of size (10 * 32).
Figure 5 shows the tweet feature matrix extraction us-
ing transformer-based architecture.
4.3 User Feature Matrix Extraction for
Measuring User Credibility
There are various features of the user that are read-
ily available on the user profile and can be used to
measure the credibility of the user posting the news.
We have used the following features to describe each
user: Follower’s count, Friends count, Statuses count,
Account verified or not and Location mentioned or
not. A user feature vector can be obtained by using
the normalized values of above measures. We then
stack the set of user feature vectors of all tweets cor-
responding to a news article to obtain a user feature
matrix for that article. So, if a user feature vector is
of size 32, then a user feature matrix will be of size
Figure 5: Tweet Feature Matrix using Transformer-based
Architecture.
Figure 6: Extracting User Feature Vector.
(10, 32) because we have considered 10 tweets corre-
sponding to a news article. Figure 6 shows the user
feature extraction model and Figure 7 shows the com-
plete architectural view.
4.4 Social Context Feature Matrix
We have obtained both tweet response (10, 32) and
user feature matrix (10, 32). Now to obtain social
context feature matrix we will just concatenate the
above matrices and the size of the matrix obtained
will be (10, 64).
4.4.1 Analysis of Social Context Feature Matrix
We need to capture both local and global variations
in the social context data. Global variations can be
captured using self-attention like mechanisms as they
analyse the entire set of tweet responses and user
characteristics for a particular news and select glob-
ally which are the prominent ones for classifying the
news correctly. To capture local features, time series
analysis of the social context data can be done using
RNNs. They analyse the variation in the social con-
text data as time progresses. We have used GRUs for
Impact of Transformer-Based Models and User Clustering in Early Fake News Detection in Social Media
893

this purpose. Later both global and local variations
can be concatenated to obtain the final social context
vector for the news article.
4.4.2 Self-Attention Mechanism (Capturing
Global Variations)
Given a sequence of K tweets (tweet response + user
characteristics) which is represented by social con-
text feature matrix (K * 64), not all of them have
the same ability to discriminate true and fake news.
Some special text response generated by some spe-
cial type of user may reflect the truthfulness of a con-
cerned news article more significantly, thus should be
somehow highlighted in the entire propagation path.
Thus, our detection model should learn how much
attention should be given to each tweet. The self-
attention module will multiply each tweet vector with
an attention score between 0 and 1. Hence all the rele-
vant tweet vectors will be multiplied with an attention
score close to 1 and all irrelevant tweets will be mul-
tiplied by scores close to 0. These attention scores
will be calculated using weight matrix which will be
trained along with the model. The weighted sum of
all these tweet vectors will form global social context
vector.
4.4.3 Time Series Analysis Using Gated
Recurrent Units (Capturing Local
Variations)
We have a sequence of K tweets <
(x
1
,t
1
), . . . ., (x
k
,t
k
) > where x
j
is the vector
representing concatenation of user characteristics and
tweet response and t
j
is the time of posting of tweet.
Now we will feed this sequence of tweets to GRU for
obtaining hidden state at each time step which will
later be used to obtain local social context vector.
4.4.4 GRU Based Local Feature Extraction
For the t
t
h social context vector in the sequence i.e.,
x
t
, a GRU takes in input as x
t
, h
(
t − 1) and produces
h
t
as output according to the following formulas:
z
t
= σ(U
z
x
t
+W
z
h
t−1
) (1)
r
t
= σ(U
r
x
t
+W
r
h
t−1
) (2)
e
h
t
= tanh(U
h
x
t
+ h
t−1
⊙ w
h
r
t
) (3)
h
t
= (1 − z
t
) ⊙ h
t−1
+ z
t
⊙
e
h
t
(4)
We then apply mean pooling to reduce the sequence
of output vectors < h
1
, . . . ., h
k
> produced by GRU
units into a single vector which is the average of the
above vectors produced at each time step. This vector
obtained is the local social context vector.
Figure 7: Complete Model Architecture View.
4.4.5 Concatenation of both Representations
(Local and Global) of Social Context
Vector
Both the representations can be concatenated into a
single vector that represents the final social context
vector. It can then be concatenated with content fea-
ture vector to obtain final feature vector which can be
fed into a multi-layer feedforward neural network that
predicts the class label for the news.
4.5 User Clustering
Social context features usually includes user-based
features and text-based features. Apart from these
two, it also captures structure-based features which
involves the relationships among the users that are
involved. In the practical scenario, with the limited
availability of open dataset that includes the user re-
lationships, even from popular microblogging web-
sites like twitter, finding an alternative way to capture
the structure-based features is necessary. In this re-
gard, we would like to extend the base architecture
with user clustering. Usually, there are two major cat-
egories of clustering:
• Hard Clustering. Where each data point has a
fixed cluster label.
• Soft Clustering. A data point can co-exist in mul-
tiple clusters with certain probability.
Extended the framework to implement K-means
which is hard clustering method and Fuzzy C-means
clustering algorithms which falls under soft cluster-
ing. User behaviour is not always binary, few of the
users tend to spread both fake news as well as real
news (knowingly or unknowingly). Having the fuzzi-
ness gives us the flexibility to have a control over the
cluster membership thresholds, which is more prac-
tical in the real world. To understand the clustering
ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods
894

Figure 8: Model Architecture with User Clustering.
methods that are in discussion in brief, K-means clus-
tering initialize the centers, assign memberships to
the datapoints and recompute the centers and repeat
until it converges. Whereas, Fuzzy C-Means Clus-
tering initialise the memberships randomly first, up-
dates the class centers using these memberships, com-
putes Euclidean distance of samples from centers, fi-
nally re-update the memberships and repeat until it
converges. Once we obtain the cluster memberships,
cluster weights are given to each of the cluster and its
members based on the relative size of the cluster they
are in. Now this cluster weights are used to scale the
feature vectors of corresponding users. As users with
huge friendship/follower network can be an outlier
and could possibly impact the model via its feature
vectors. To reduce the impact of outliers, the cluster
weights are user to scale the feature vectors of the re-
spective users. Figure 8 shows the model architecture
with User Clustering.
5 EXPERIMENTS AND RESULTS
The FakenewsNet repository was used for the exper-
imental analysis. The labeled data has been fact-
checked by PolitiFact and GossipCop. Table 2 shows
the accuracy and F1 score of the PolitiFact data. Table
3 shows the accuracy and F1 score of the GossipCop
data. Table 4 shows the Accuracy and F1 score re-
lated to various content based and content-social con-
text ensemble based models.
Table 2: Accuracy and F1-score on PolitiFact Data.
PolitiFact Content
(Glove
based)+
Social con-
text(global)
Content
(Glove
based) +
social con-
text (local
+global)
Content (
transform-
ers) + social
context
(global)
Content (
transform-
ers) + social
context
(local +
global)
Validation
Accuracy
85.5% 86.18% 83.22% 84.12%
F1-Score 84.33% 85.25% 83.56% 82.22%
Considering local and global social context-based
features with the content-based features improved the
Table 3: Accuracy and F1-score on GossipCop Data.
GossipCop Content
(glove) +
social con-
text (global)
Content (
transform-
ers) + social
context
(global)
Content (
transform-
ers) + social
context
(global)
Content (
transform-
ers) + social
context
(local +
global)
Validation
Accuracy
87.2% 88.4% 83.22% 84.12%
F1-Score 86.7% 87.3% 83.56% 82.22%
performance by significant margin. And since we
have used only a limited no. of tweets, our model will
be able to detect news at an early stage of news prop-
agation. The results of our model are comparable to
the results of models that uses entire propagation net-
work which takes months to build and hence violates
the constraint of early detection of fake news.
We can observe that sentence level models, sen-
tence level CNN and sentence level Roberta, Sen-
tence Level CNN with Fuzzy clustering performed
better than their counterparts which considered whole
article as a single sentence/entity as shown in Table
4. Also, adding social context-based features to the
content-based features improved the performance by
significant margin. And since we have used only a
limited no. of tweets, our model will be able to detect
news at an early stage of news propagation. The re-
sults of our model are comparable to the results of
models that uses entire propagation network which
takes months to build and hence violates the con-
straint of early detection of fake news.
6 CONCLUSION AND FUTURE
WORK
Even using only about 20 % social context data, we
were able to achieve accuracy comparable to models
that uses entire social context data hence violate the
constraint of early detection. Ensembling the content-
based model with social context based is the way
to deal with generalizability issue of content-based
model. Using both CNN and RNN ( Time Series
Analysis) we can capture both local and global varia-
tions in the social context data. Cluster credibility and
self-attention helps to identify which tweet and user
should be taken into consideration to classify news as
fake or real. For the future work retweets can be con-
sidered with the tweets which were obtained in the
same time window. Comments on tweets and retweets
can also be considered. After including retweets, we
can build a social context graph for the propagation of
news and instead of using text CNN or sentence level
CNN we can use graph CNN for feature extraction
and news classification. Introducing fuzzy nature to
the clustering certainly improved the performance of
Impact of Transformer-Based Models and User Clustering in Early Fake News Detection in Social Media
895

Table 4: Accuracy and F1 score related to various content based and content-social context ensemble based models.
Base
Model
Content-Based Content-Social Context Ensemble
Evalu-
ation
Metrics
Text-
CNN
Sentence
Level
CNN
RoBERTa Sentence
- Level
RoBERTa
Sentence-
CNN+
Social
Context
Based
Model
Sentence
RoBERTa +
Social Con-
text Based
Model
Clustering Methods
Sentence
CNN+K-
Means
Sentence
CNN+
Fuzzy
C-
Means
Accuracy 73.3% 80.5% 75.4% 78.6% 86.4% 83.3% 80.2% 83.6%
F1-
Score
0.67% 0.77% 0.73% 0.76% 0.85% 0.82% 0.69% 0.84%
the model. As Fuzzy C-means clustering only consid-
ers Euclidian distance to compute, using a metric like
Mahalanobis distance which also captures the spatial
metrics can improve the model.
REFERENCES
Afroz, S., Brennan, M., and Greenstadt, R. (2012). De-
tecting hoaxes, frauds, and deception in writing style
online. In 2012 IEEE Symposium on Security and Pri-
vacy, pages 461–475. IEEE.
Castillo, C., Mendoza, M., and Poblete, B. (2011). Informa-
tion credibility on twitter. In Proceedings of the 20th
international conference on World wide web, pages
675–684.
Gupta, A., Kumaraguru, P., Castillo, C., and Meier, P.
(2014). Tweetcred: Real-time credibility assessment
of content on twitter. In International conference on
social informatics, pages 228–243. Springer.
Jin, Z., Cao, J., Zhang, Y., Zhou, J., and Tian, Q. (2016).
Novel visual and statistical image features for mi-
croblogs news verification. IEEE transactions on mul-
timedia, 19(3):598–608.
Klyuev, V. (2018). Fake news filtering: Semantic ap-
proaches. In 2018 7th International Conference on
Reliability, Infocom Technologies and Optimization
(Trends and Future Directions)(ICRITO), pages 9–15.
IEEE.
Kshetri, N. and Voas, J. (2017). The economics of “fake
news”. IT Professional, 19(6):8–12.
Liu, Y., Jin, X., Shen, H., and Cheng, X. (2017). Do rumors
diffuse differently from non-rumors? a systematically
empirical analysis in sina weibo for rumor identifica-
tion. In Pacific-Asia conference on knowledge discov-
ery and data mining, pages 407–420. Springer.
Ma, J., Gao, W., Mitra, P., Kwon, S., Jansen, B. J., Wong,
K.-F., and Cha, M. (2016). Detecting rumors from
microblogs with recurrent neural networks.
Ma, J., Gao, W., Wei, Z., Lu, Y., and Wong, K.-F. (2015).
Detect rumors using time series of social context in-
formation on microblogging websites. In Proceedings
of the 24th ACM international on conference on in-
formation and knowledge management, pages 1751–
1754.
Shu, K., Mahudeswaran, D., Wang, S., Lee, D., and Liu,
H. (2020). Fakenewsnet: A data repository with news
content, social context, and spatiotemporal informa-
tion for studying fake news on social media. Big data,
8(3):171–188.
Sun, S., Liu, H., He, J., and Du, X. (2013). Detecting event
rumors on sina weibo automatically. In Asia-Pacific
web conference, pages 120–131. Springer.
Vosoughi, S., Roy, D., and Aral, S. (2018). The spread of
true and false news online. science, 359(6380):1146–
1151.
Yang, F., Liu, Y., Yu, X., and Yang, M. (2012). Automatic
detection of rumor on sina weibo. In Proceedings of
the ACM SIGKDD workshop on mining data seman-
tics, pages 1–7.
Yang, Y., Niu, K., and He, Z. (2015). Exploiting the
topology property of social network for rumor detec-
tion. In 2015 12th International Joint Conference
on Computer Science and Software Engineering (JC-
SSE), pages 41–46. IEEE.
ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods
896