The Good, the Bad and the Bait: Detecting and Characterizing Clickbait on YouTube

Abstract

The use of deceptive techniques in user-generated video portals is ubiquitous. Unscrupulous uploaders deliberately mislabel video descriptors aiming at increasing their views and subsequently their ad revenue. This problem, usually referred to as "clickbait," may severely undermine user experience. In this work, we study the clickbait problem on YouTube by collecting metadata for 206k videos. To address it, we devise a deep learning model based on variational autoencoders that supports the diverse modalities of data that videos include. The proposed model relies on a limited amount of manually labeled data to classify a large corpus of unlabeled data. Our evaluation indicates that the proposed model offers improved performance when compared to other conventional models. Our analysis of the collected data indicates that YouTube recommendation engine does not take into account clickbait. Thus, it is susceptible to recommending misleading videos to users.

Full text

The Good, the Bad and the Bait: Detecting and
Characterizing Clickbait on YouTube
Savvas Zannettou, Sotirios Chatzis, Kostantinos Papadamou, Michael Sirivianos
Department of Electrical Engineering, Computer Engineering and Informatics
Cyprus University of Technology
Limassol, Cyprus
sa.zannettou@edu.cut.ac.cy, {sotirios.chatzis,michael.sirivianos}@cut.ac.cy, ck.papadamou@edu.cut.ac.cy
Abstract—The use of deceptive techniques in user-generated
video portals is ubiquitous. Unscrupulous uploaders deliberately
mislabel video descriptors aiming at increasing their views and
subsequently their ad revenue. This problem, usually referred to
as "clickbait," may severely undermine user experience. In this
work, we study the clickbait problem on YouTube by collecting
metadata for 206k videos. To address it, we devise a deep
learning model based on variational autoencoders that supports
the diverse modalities of data that videos include. The proposed
model relies on a limited amount of manually labeled data to
classify a large corpus of unlabeled data. Our evaluation indicates
that the proposed model offers improved performance when
compared to other conventional models. Our analysis of the
collected data indicates that YouTube recommendation engine
does not take into account clickbait. Thus, it is susceptible to
recommending misleading videos to users.
I. INTRODUCTION
Recently, YouTube surpassed cable TV in terms of popu-
larity within teenagers [28]. This is because YouTube offers
a vast amount of videos, which are always available on
demand. However, because videos are generated by the users
of the platform, known as YouTubers, a plethora of them
are of dubious quality. The ultimate goal of YouTubers is to
increase their ad revenue by ensuring that their content will get
viewed by millions of users. Several YouTubers deliberately
employ techniques that aim to deceive viewers into clicking
their videos. These techniques include: (i) use of eye-catching
thumbnails, such as depictions of abnormal stuff or attractive
adults, which are often irrelevant to video content; (ii) use of
headlines that aim to intrigue the viewers; and (iii) encapsulate
false information to either the headline, the thumbnail or the
video content. We refer to videos that employ such techniques
as clickbaits. The continuous exposure of users to clickbaits
cause frustration and degraded user experience ( see Fig. 1 ) .
The clickbait problem is essentially a peculiar form of the
well-known spam problem [38], [37], [36], [17], [20]. In spam,
malicious users try to deceive users by sending them mislead-
ing messages mainly to advertise websites or perform attacks
(e.g., phishing) by redirecting users to malicious websites.
Nowadays, the spam problem is not as prevalent as a few
years ago due to the deployment of systems that diminish
it. Furthermore, users have an increased awareness of typical
spam content (e.g., emails, etc.) and they can effortlessly
discern it. However, this is not the case for clickbait, which
Fig. 1: Comments that were found in clickbait videos. The users’
frustration is apparent (we omit users’ names for ethical reasons).
usually contains hidden false or ambiguous information that
users or systems might not be able to perceive.
Recently, the aggravation of the fake news problem has
induced broader public attention to the clickbait problem.
For instance, Facebook aims at removing clickbaits from its
newsfeed [27], [14]. In this work, we focus on YouTube
for various reasons: i) anecdotal evidence suggests that the
problem exists in YouTube [8] and ii) to the best of our
knowledge, YouTube relies on users to flag suspicious videos
and then manually review them. To this extent, this approach
is deemed to be inefficient. Hence, the need for an automated
approach that minimizes human intervention is indisputable.
To attain this goal, we leverage some recent advances in the
field of Deep Learning [22] by devising a novel formulation of
variational autoencoders (VAEs) [21], [23] that fuses different
modalities of a YouTube video, and infers latent correlations
between them.
The proposed model infers a latent variable vector for
each video that encodes a high-level representation of the
content and the correlations between the various modalities.
The significance of learning to compute such concise repre-
sentations is that: (i) this learning procedure can be robustly
performed by leveraging large unlabeled data corpora; and (ii)
the obtained representations can be subsequently utilized to
drive the classification process, with very limited requirements
in labeled data.
To this end, we formulate the encoder part of the devised
VAE model as a 2-component finite mixture model [10]. That
is, we consider a set of alternative encoder models that may
generate the data pertaining to each video. The decision of
which specific encoder (corresponding to one possible class)
generates each is obtained via a trainable probabilistic gating
network [29]; this constitutes an integral part of the developed

autoencoder. The whole model is trained in an end-to-end
fashion, using the available training data, both the unlabeled
and the few labeled ones. The latter are specifically useful
for appropriately fitting the postulated gating network that
infers the posterior distribution of mixture component (and
corresponding class) allocation.
Contributions. We propose a deep generative model that
allows for combining data from as diverse modalities as video
headline text, thumbnail image and tags text, as well as various
numerical statistics, including statistics from comments. Most
importantly, the proposed model allows for successfully ad-
dressing the problem of learning from limited labeled samples
and numerous unlabeled ones (semi-supervised learning). This
is achieved by postulating a deep variational autoencoder that
employs a finite mixture model as its encoder. In this context,
mixture component assignment is regulated via an appropriate
gating network; this also constitutes the eventually obtained
classification mechanism of our deep learning system. We
provide a large scale analysis on YouTube; we show that, with
respect to the collected data, its recommendation engine does
not consider how misleading a recommended video is. Hence,
it is susceptible to recommending clickbait videos to its users.
II. METHODOLOGY
By leveraging YouTube’s Data API, between August and
November of 2016, we collect metadata of videos published
between 2005 and 2016. Specifically, we collected the fol-
lowing data descriptors for 206k videos: (i) basic details like
headline, tags, etc.; (ii) thumbnail; (iii) comments from users;
(iv) statistics (e.g., views, likes, etc.); and (v) related videos
based on YouTube’s recommendation system. We started our
retrieval from a popular (108M views) clickbait video [1]
and iteratively collected all the related videos as were rec-
ommended by YouTube. Note that this approach enables us
to study interesting aspects of the problem, by constructing a
graph that captures the relations (recommendations) between
videos.
To get labeled data, we opted for two different approaches.
First, we manually reviewed a small subset of the collected
data by inspecting the headline, the thumbnail, comments from
users, and video content. Specifically, we watched the whole
video and compared it to the thumbnail and headline. A video
is considered clickbait only if the thumbnail and headline
deviate substantially from its content. However, this task is
both cumbersome and time consuming; thus, we elected to
retrieve more data that are labeled. To this end, we compiled
a list of channels (available at [2]) that habitually employ
clickbait techniques and channels that do not. To obtain the list
of channels, we used a pragmatic approach; we found channels
that are outed by other users as clickbait channels. For each
channel, we retrieved up-to 500 videos, hence creating a larger
labeled dataset. The overall labeled dataset consists of (i)
1,071 clickbaits and 955 non-clickbaits obtained from the
manual review process and (ii) 8,999 clickbaits and 8,470 non-
clickbaits obtained from the distinguished list of channels. The
importance of this dataset is two-fold, as it allow us to study
the problem and is instrumental for training our deep learning
model.
A. Manually Reviewed Ground Truth Dataset Analysis
In order to better grasp the problem, we perform a compar-
ative analysis of the manually reviewed ground truth.
Category. Table 1 reports the categories we find on the videos.
In total, we find 15 categories but we only show the top
five in terms of count for brevity. We observe that most
clickbaits exist in the Entertainment and Comedy categories,
whereas non-clickbaits are prevalent in the Sports category.
This indicates that, within this dataset, YouTubers employ
clickbait techniques on videos for entertainment.
Headline. YouTubers normally employ deceptive techniques
on the headline like the use of exaggerating phrases. To verify
that this applies to our ground truth dataset, we perform
stemming to the words that are found in clickbait and non-
clickbait headlines. Fig. 2 (a) depicts the ratio of the top 20
stems that are found in our ground truth clickbait videos (i.e.,
95% of the videos that contain the stem “sexi” are clickbait).
In essence, we observe that magnetizing stems like “sexi” and
“hot” are frequently used in clickbait videos, whereas their
use in non-clickbaits is low. The same applies to words used
for exaggeration, like “viral” and “epic”.
Thumbnail. To study the thumbnails, we make use of
Imagga [19], which offers descriptive tags for an image. We
perform tagging of all the thumbnails in our ground truth
dataset. Fig. 2(b) demonstrates the ratio of the top 10 Imagga
tags that are found in the manually reviewed ground truth. We
observe that clickbait videos typically use sexually-appealing
thumbnails in their videos in order to attract viewers. For
instance, 81% of the videos’ thumbnail of which contains the
“pretty” tag are clickbaits.
Tags. Tags are words that are defined by YouTubers before
publishing and can dictate whether a video will emerge on
users’ search queries. We notice that clickbaits use specific
words on tags, whereas non-clickbaits do not. Fig. 2 (c) depicts
the ratio of the top 20 stems that are found in clickbaits.
We observe that many clickbait videos use tags like “try not
to laugh”, “viral”, “hot” and “impossible”; phrases that are
usually used for exaggeration.
Statistics. Fig. 2 (d) shows the normalized score of the video
statistics for both classes of videos. Interestingly, clickbaits
and non-clickbaits videos have similar views; suggesting that
viewers are not able to easily discern clickbait videos, hence
clicking on them. Also, non-clickbait videos have more likes
and less dislikes than clickbaits. This is reasonable as many
users feel frustrated after watching clickbaits.
Comments. We notice that users on YouTube implicitly flag
suspicious videos by commenting on them. For instance,
we note several comments like the following: “the title is
misleading”, “i clicked because of the thumbnail”, “where
is the thumbnail?” and “clickbait”. Hence, we argue that
comments from viewers is a valuable resource for assessing
videos. To this end, we analyze the ground truth dataset to
extract the mean number of occurrences of words widely used

(a) (b) (c)
(d) (e)
Category Clickbaits (%) Non-clickbaits (%)
Entertainment 406 (38%) 308 (32%)
Comedy 318 (29%) 228 (24%)
People & Blogs 155 (14%) 115 (12%)
Autos & Vehicles 33 (3%) 49 (5%)
Sports 29 (3%) 114 (12%)
Fig. 2 & TABLE I: Analysis of the manually reviewed ground truth dataset. Normalized mean scores for: (a) stems from headline text; (b)
tags derived from thumbnails; (c) stems from tags that were defined by uploaders; (d) video statistics; and (e) comments that contain words
for flagging suspicious videos. Table 1 shows the top five categories (and their respective percentages) in our ground truth dataset.
Source Destination Norm. Mean
clickbait clickbait 4.1
clickbait non-clickbait 2.73
non-clickbait clickbait 2.75
non-clickbait non-clickbait 3.57
TABLE II: Normalized mean of related videos for clickbait and
non-clickbait videos in the ground truth dataset
for flagging clickbait videos. Fig. 2 (e) depicts the normalized
mean scores for the identified words. We observe that these
words were greatly used in clickbait comments but not in non-
clickbaits. Also, it is particularly interesting that comments
referring to the video’s thumbnail were found 2.5 times more
often in clickbait than in non-clickbaits.
Graph Analysis. Users often watch videos according to
YouTube’s recommendations. From manual inspections, we
have noted that when watching a clickbait video, YouTube is
more likely to recommend another clickbait video. To confirm
this against our data, we create a directed graph G= (V,E),
where Vthe videos and Ethe connections between videos
pointing to one another via a recommendation. Then, for all the
videos, we select their immediate neighbors in the graph, and
count the videos that are clickbaits and non-clickbaits. Table II
depicts the normalized mean of the number of connected
videos for each class. We apply a normalization factor to
mitigate the bias towards clickbaits, which have a slightly
greater number in our ground truth. We observe that, when
a user watches a clickbait video, they are recommended 4.1
clickbait videos on average, as opposed to 2.73 non-clickbait
recommendations. A similar pattern holds for non-clickbaits;
a user is less likely to be served a clickbait when watching a
non-clickbait.
YouTube’s countermeasures. To get an insight on whether
YouTube employs any countermeasures, we calculate the num-
ber of offline (either deleted by YouTube or removed by the
uploader) videos in our manually reviewed ground truth, as
of January 10, 2017 and April 30, 2017. We found that only
3% (January 10th) and 10% (April 30th) of the clickbaits are
offline. Similarly, only 1% (January 10th) and 5% (April 30th)
of the non-clickbaits are offline. To verify that the ground
truth dataset does not consist of only recent videos (thus, just
published and not yet detected) we calculate the mean number
of days that passed from the publication date up to January 10,
2017. We find that the mean number of days for the clickbaits
is 700, while it is 917 days for the non-clickbaits. The very
low offline ratio, as well as the high mean number of days,
indicate that YouTube is not able to tackle the problem in a
timely manner.
III. CLICKBAIT DETE CT IO N MOD EL
A. Processed Modalities
Our model processes the following modalities: (i) Headline:
For the headline, we consider both the content and the style
of the text. For the content of the headline, we use sent2vec
embeddings [26] trained on Twitter data. For the style of the
text, we use the features proposed in [6]; (ii) Thumbnail: We
scale down the images to 28x28 and convert them to grayscale.
This way, we decrease the number of trainable parameters
for our developed deep network, thus speeding training time
up without compromising achievable performance; (iii) Com-
ments: We preprocess user comments to find the number of
occurrences of words used for flagging videos. We consider
the following words: “misleading, bait, thumbnail, clickbait,
deceptive, deceiving, clicked, flagged, title”; (iv) Tags: We
encode the tags’ text as a binary representation of the top 100

Fig. 3: Overview of the proposed model. The dotted rectangle
represents the encoding component of our model.
words that are found in the whole corpus; and (v) Statistics
(e.g., likes, views, etc.).
B. Model Formulation
In Fig. 3, we provide an overview of the proposed model.
The thumbnail is initially processed, at the encoding part of
the proposed model, by a CNN [40]. We use a CNN that
comprises four convolutional layers, with 64 filters each, and
ReLU activations. The first three of these layers are followed
by max-pooling layers. The fourth is followed by a simple
densely connected layer, which comprises 32 units with ReLU
activations. This initial processing stage allows for learning to
extract a high-level, 32-dimensional vector of the thumbnail,
which contains the most useful bits of information for driving
classification.
The overarching goal of the devised model is to limit
the required availability of labeled data, while making the
most out of large corpora of (unlabeled) examples. To this
end, after this first processing stage, we split the encoding
part into two distinct subencoders that work in tandem. Both
these subencoders are presented with the aforementioned 32-
dimensional thumbnail representation, fused with the data
stemming from all the other available modalities. This results
in a 855-dimensional input vector, first processed by one dense
layer network (Fusing Network) that comprises 300 ReLU
units. Due to its large number of parameters, this dense layer
network may become prone to overfitting; hence, we regularize
using the prominent Dropout technique [34]. We use a Dropout
level of d= 0.5; this means that, at each iteration of the
training algorithm, 50% of the units are randomly omitted
from updating their associated parameters. The obtained 300-
dimensional vector, say h(), is the one eventually presented to
both postulated subencoders.
The rationale behind this novel configuration is motivated
by a key observation; the two modeled classes are expected
to entail significantly different patterns of correlations and
latent underlying dynamics between the modalities. Hence,
it is plausible that each class can be adequately and effec-
tively modeled by means of distinct, and different, encoder
distributions, inferred by the two subencoders. Each of these
subencoders are dense-layer networks comprising a hidden
layer with 20 ReLU units, and an output layer with 10 units.
Since the devised model constitutes a VAE, the output units
of the subencoders are of a stochastic nature; specifically,
we consider stochastic outputs, say ˜
zand ˆ
z, with Gaussian
(posterior) densities, as usual in the literature of VAEs [21],
[23]. Hence, what the postulated subencoders actually compute
are the means, ˜
µand ˆ
µ, as well as the (diagonal) covariance
matrices, ˜
σ2and ˆ
σ2, of these Gaussian posteriors. On this
basis, the actual subencoder output vectors, ˜
zand ˆ
z, are
sampled each time from the corresponding (inferred) Gaussian
posteriors. Note that our modeling selection of sharing the ini-
tial CNN-based processing part between the two subencoders
allows for significantly reducing the number of trainable
parameters, without limiting the eventually obtained modeling
power.
Under this mixture model formulation, we need to establish
an effective mechanism for inferring which observations (i.e.,
videos) are more likely to match the learned distribution of
each component subencoder. This is crucial for effectively
selecting between the samples of ˜
zor ˆ
zat the output of the
encoding stage of the devised model. In layman terms, this
can be considered to be analogous to a (soft) classification
mechanism. This mechanism can be obtained by computation
of the posterior distribution of mixture component membership
of each video (also known as "responsibility" in the literature
of finite mixture models [25]). To allow for inferring this
posterior distribution, in this work we postulate a gating
network. This is a dense-layer network, which comprises one
hidden layer with 100 ReLU units, and is presented with
the same vector, h(), as the two postulated subencoders. It
is trained alongside the rest of the model, and it is the only
part of the model that requires availability of labeled data for
its training.
Note that this gating network entails only a modest number
of trainable parameters, since both the size of its input as
well as of its single hidden layer are rather small. As such,
it can be effectively trained even with limited availability of
labeled data. This is a key merit of our approach, which fully
differentiates it from conventional classifiers that are presented
with raw observed data, which typically are prohibitively high-
dimensional.
To conclude the formulation of the proposed VAE, we
need to postulate an appropriate decoder distribution, and a
corresponding network that infers it. In this work, we opt
for a simple dense-layer neural network, which is fed with
the (sampled) output of the postulated finite mixture model
encoder, and attempts to reconstruct the original modalities.
Specifically, we postulate a network comprising one hidden

layer with 300 ReLU units, and a set of 823 output units,
that attempt to reconstruct the original modalities, with the
exception of the thumbnail. The reason why we ignore the
thumbnail modality from the decoding process is the need
of utilizing deconvolutional network layers to appropriately
handle it, which is quite complicated. Hence, we essentially
treat the thumbnail modality as side-information, that regulates
the inferred posterior distribution of the latent variables z
(encoder) in the sense of a covariate, instead of an observed
modality in the conventional sense. It is empirically known
that such a setup, if appropriately implemented, does not
undermine modeling effectiveness [30].
Let us denote as xnthe set of observable data per-
taining to the nth available video. We denote as xn=
{xhe
n,xth
n,xta
n,xco
n,xst
n}the set of five distinct modalities,
i.e., headline, thumbnail, tags, comments, and statistics, re-
spectively. Then, based on the above description, the encoder
distribution of the postulated model reads:
q(zn|xn) =q(˜
zn|xn)q(cn=1|xn)q(ˆ
zn|xn)q(cn=0|xn)(1)
Here, znis the output of the encoding stage of the proposed
model that corresponds to xn,˜
znis the output of the first
subencoder, corresponding to the clickbait class, ˆ
znis the
output of the second subencoder, corresponding to the non-
clickbait class, and cnis a latent variable indicator of whether
xnbelongs to the clickbait class or not. We also postulate
q(˜
zn|xn) = N(˜
zn|˜
µ(xn;˜
θ),diag ˜
σ2(xn;˜
θ)) (2)
q(ˆ
zn|xn) = N(ˆ
zn|ˆ
µ(xn;ˆ
θ),diag ˆ
σ2(xn;ˆ
θ)) (3)
Here, the ˜
µ(xn;˜
θ)and ˜
σ2(xn;˜
θ)are outputs of a deep
neural network, with parameters set ˜
θ, that corresponds to the
clickbait class subencoder; it comprises a first CNN-type part
that processes the observed thumbnails, and a further densely-
connected network part that fuses and processes the rest of
the observed modalities, as described previously. Similarly, the
ˆ
µ(xn;ˆ
θ)and ˆ
σ2(xn;ˆ
θ)are outputs of a deep neural network
with parameters set ˆ
θ, that corresponds to the non-clickbait
class subencoder.
The posterior distribution of mixture component allocation,
q(cn|xn), which is parameterized by the aforementioned gat-
ing network, is a simple Bernoulli distribution that reads
q(cn|xn) = Bernoulli($(h(xn); ϕ)) (4)
Here, $(h(xn); ϕ)∈[0,1] is the output of the gating network,
with trainable parameters set ϕ. This is presented with an
intermediate encoding of the input modalities (shared with
the subencoder networks), h(xn), as described previously, and
infers the probability of xnbelonging to the clickbait class.
Lastly, the postulated decoder distribution reads
p(xn|zn) = N(xhe
n,xta
n,xco
n,xst
n|µ(zn;φ),diag σ2(zn;φ))
(5)
where the means and diagonal covariances, µ(zn;φ)and
σ2(zn;φ), are outputs of a deep network with trainable
parameters set φ, configured as described previously.
C. Model Training
Let us consider a training dataset X={xn}N
n=1 that
consists of Nvideo samples. A small subset, Xl, of size Mof
these samples is considered to be labeled, with corresponding
labels set Y={ym}M
m=1. Then, following the VAE literature
[21], model training is performed by maximizing the evidence
lower bound (ELBO) of the model over the parameters set
{˜
θ,ˆ
θ,ϕ,φ}. The ELBO of our model reads:
log p(X)≥L(˜
θ,ˆ
θ,ϕ,φ|X) = −
N
X
n=1
KLq(zn|xn)||p(zn)
+γ
N
X
n=1
E[log p(xn|zn)] + X
xm∈Xl
log q(cm=ym|xm)
(6)
Here, KLq||pis the KL divergence between the distribution
q(·)and the distribution p(·), while E[·]is the (posterior)
expectation of a function w.r.t. its entailed random (latent)
variables. Note also that, in the ELBO expression (6), the
introduced hyperparameter γis a simple regularization con-
stant, employed to ameliorate the overfitting tendency of
the postulated decoder networks, p(xn|zn). We have noticed
that this simple trick yields a significant improvement in
generalization capacity.
In Eq. (6), the posterior expectation of the log-likelihood
term p(xn|zn)cannot be computed analytically, due to the
nonlinear form of the decoder. Hence, we must approximate
it by drawing Monte-Carlo (MC) samples from the posterior
(encoder) distributions (2)-(3). However, MC gradients are
well-known to suffer from high variance. To resolve this
issue, we utilize a smart re-parameterization of the drawn MC
samples. Specifically, following the related derivations in [21],
we express these samples in the form of a differentiable
transformation of an (auxiliary) random noise variable ; this
random variable is the one we actually draw MC samples from:
˜
z(s)
n=˜
µn+˜
σn·(s)
n,(s)
n∼ N (0,I)(7)
ˆ
z(s)
n=ˆ
µn+ˆ
σn·(s)
n(8)
Hence, such a re-parameterization reduces the computed ex-
pectations into averages over samples from a random variable
with low (unitary) variance, . This way, by maximizing
the obtained ELBO expression, one can yield low-variance
estimators of the sought (trainable) parameters, under some
mild conditions [21]. Turning to the maximization process of
L(˜
θ,ˆ
θ,ϕ,φ|X), this can be effected using modern stochastic
optimizations, such as AdaGrad [16] or RmsProp [39].
IV. EXP ER IM EN TAL EVALUATIO N
A. Experimental Setup
Our model is implemented with Keras [12] and Tensor-
Flow [24].
MC Samples. To perform model training, we used S= 10
drawn MC samples, (s); we found that increasing this value
does not yield any statistically significant accuracy improve-
ment, despite the associated increase in computational costs.

Initialization. We employ Glorot Uniform initialization [18].
This scheme allow us to train deep network architectures
without the need of layerwise pretraining. This is effected by
initializing the network weights in a way which ensures that
the signal propagated across the network layers remains in a
reasonable range of values, irrespectively of the network depth
(i.e., it does not explode to infinity or vanish to zero).
Stochastic optimization. We found that RmsProp works better
than the AdaGrad algorithm suggested in [21]. The RmsProp
hyperparameter values we used comprise an initial learning
rate equal to 0.001, ρ= 0.9, and ε= 10−8.
Prediction generation. To predict the class of a video,
xn, we compute the mixture assignment posterior distribu-
tion q(cn|xn), inferred via the postulated gating network
$(h(xn); ϕ). On this basis, assignment is performed to the
clickbait class if $(h(xn); ϕ)>0.5.
Baselines. To evaluate our model, we compare it against
two baseline models. First, a simple Support Vector Machine
(SVM) with parameters γ= 0.001 and C= 100. Second, a
supervised deep network (SDN) that comprises (i) the same
CNN as the proposed model; and (ii) a 2-layer fully-connected
neural network with Dropout level of d= 0.5.
Evaluation. We train our proposed model using the entirety of
the available unlabeled dataset, as well as a randomly selected
80% of the available labeled dataset, comprising an equal
number of clickbait and non-clickbait examples. Subsequently,
the trained model is used to perform out-of-sample evaluation;
that is, we compute the classification performance of our
approach on the fraction of the available labeled examples
that were not used for model training.
Model Accuracy Precision Recall F1 Score
SVM 0.882 0.909 0.884 0.896
SDN 0.908 0.920 0.907 0.917
Proposed Model (U = 25% ) 0.915 0.918 0.926 0.923
Proposed Model (U = 50% ) 0.918 0.918 0.934 0.926
Proposed Model (U = 100% ) 0.924 0.921 0.942 0.931
TABLE III: Performance metrics for the evaluated methods. We also
report the performance of our model when using only 25% or 50%
of the available unlabeled data.
Table III reports the performance of the proposed model as
well as the two considered baselines. We observe that neural
network-based approaches, such as a simple neural network
and the proposed model, outperform SVMs in terms of all
the considered metrics. Specifically, the best performance is
obtained by the proposed model, which outperforms SVMs
by 3.8%, 1.2%, 5.8% and 3.5% on accuracy, precision, recall,
and F1 score, respectively. Further, to assess the importance
of using unlabeled data, we also report results with reduced
unlabeled data. We observe that, using only 25% of the
available unlabeled data, the proposed model undergoes a
substantial performance decrease, as measured by all the em-
ployed performance metrics. This performance deterioration
only slightly improves when we elect to retain 50% of the
available unlabeled data.
Corpus-Level Inference Insights. Having demonstrated the
performance of our model, we now provide some insights into
the obtained inferential outcomes on the whole corpus of data.
From the 206k examples, our model predicts that 84k (41%)
of them are clickbaits whereas 122k (59%) are non-clickbaits.
The considerable percentage of clickbaits in the corpus, in
conjunction with the data collection procedure, suggests that,
with respect to the collected data, YouTube does not consider
misleading videos in their recommendations. Note also that
we have performed an analysis of the whole corpus akin to
the ground truth dataset analysis of Section II.A. The obtained
results follow the same pattern as for the ground truth dataset.
Specifically, we have found that the normalized mean values,
reported in Table II for the labeled data, for the whole corpus
become equal to 11.18, when it comes to pairs of clickbait
videos, whereas for clickbait and non-clickbaits the mean is
2.62. This validates our deduction that it is more likely to
be recommended a clickbait video when viewing a clickbait
video on YouTube.
V. RELATED WORK
The clickbait problem is also identified by prior work that
proposes tools for alleviating the problem in various web
portals. Specifically, Chen et al. [11] provide useful informa-
tion regarding the clickbait problem and future directions for
tackling the problem using SVM and Naive Bayes approaches.
Rony et al. [32] analyze 1.67M posts on Facebook in order to
understand the extent and impact of the clickbait problem as
well as users’ engagement. For detecting clickbaits, they pro-
pose the use of sub-word embeddings with a linear classifier.
Potthast et al. [31] focus on the Twitter platform where they
suggest the use of Random Forests for distinguishing tweets
that contain clickbait content. Furthermore, Chakraborty et
al. [9] propose the use of SVMs in conjunction with a browser
add-on for offering a detection system to end-users of news
articles. Moreover, Biyani et al. [6] recommend the use of
Gradient Boosted Decision Trees for clickbait detection in
news articles. They also demonstrate that the degree of infor-
mality in the content of the landing page can help in discerning
clickbait news articles. To the best of our knowledge, Anand
et al. [4] is the first work that suggests the use of deep learning
techniques for mitigating the clickbait problem. Specifically,
they propose the use of Recurrent Neural Networks in conjunc-
tion with word2vec embeddings for identifying clickbait news
articles. Agrawal [3] propose the use of CNNs in conjunction
with word2vec embeddings for discerning clickbait headlines
in Reddit, Facebook and Twitter. Other efforts include browser
add-ons [15], [13], [7] and manually associating user accounts
with clickbait content [33], [35], [5].
Remarks. In contrast to the aforementioned works, we focus
on the YouTube platform and propose a deep learning model
that: (i) successfully and properly fuse and correlate the diverse
set of modalities related to a video; and (ii) leverage large
unlabeled datasets, while imposing much limited requirements
in labeled data availability.

VI. CONCLUSION
In this work, we have explored the use of variational
autoencoders for tackling the clickbait problem on YouTube.
Our approach constitutes the first proposed semi-supervised
deep learning technique in the field of clickbait detection. This
way, it enables more effective automated detection of clickbait
videos in the absence of large-scale labeled data. Our analysis
indicates that YouTube recommendation engine does not take
into account the clickbait problem in its recommendations.
ACK NOW LE DG ME NT S
This project has received funding from the European
Union’s Horizon 2020 Research and Innovation program,
under the Marie Skłodowska-Curie ENCASE project (Grant
Agreement No. 691025). We also gratefully acknowledge the
support of NVIDIA Corporation, with the donation of the Titan
Xp GPU used for this research.
REFERENCES
[1] Clickbait video. https://www.youtube.com/watch?v=W2WgTE9OKyg.
[2] List of ground truth channels. https://goo.gl/ZSabkn, 2018.
[3] A. Agrawal. Clickbait detection using deep learning. In NGCT, 2016.
[4] A. Anand, T. Chakraborty, and N. Park. We used Neural Networks
to Detect Clickbaits: You won’t believe what happened Next! arXiv
preprint arXiv:1612.01340, 2016.
[5] Anti-Clickbait. https://twitter.com/articlespoiler, 2015.
[6] P. Biyani, K. Tsioutsiouliklis, and J. Blackmer. "8 Amazing Secrets
for Getting More Clicks": Detecting Clickbaits in News Streams Using
Article Informality. 2016.
[7] B.s detector browser extension. http://bsdetector.tech/.
[8] R. Campbell. You Wont Believe How Clickbait is Destroying YouTube!,
2017. https://tritontimes.com/11564/columns/you-wont-believe-how-
clickbait-is-destroying-youtube/.
[9] A. Chakraborty, B. Paranjape, S. Kakarla, and N. Ganguly. Stop
Clickbait: Detecting and Preventing Clickbaits in Online News Media.
2016.
[10] S. P. Chatzis, D. I. Kosmopoulos, and T. A. Varvarigou. Signal Modeling
and Classification Using a Robust Latent Space Model Based on t
Distributions. TOSP, 2008.
[11] Y. Chen, N. J. Conroy, and V. L. Rubin. Misleading Online Content:
Recognizing Clickbait as False News. In ACM MDD, 2015.
[12] F. Chollet. Keras. https://github.com/fchollet/keras, 2015.
[13] Clickbait remover for Facebook. http://bit.ly/2DIrpj6.
[26] M. Pagliardini, P. Gupta, and M. Jaggi. Unsupervised Learning of
Sentence Embeddings using Compositional n-Gram Features. arXiv,
2017.
[14] J. Constine. Facebook feed change fights clickbait post by post in 9
more languages, 2017. http://archive.is/l8hIt.
[15] Downworthy browser extension. http://downworthy.snipe.net/.
[16] J. Duchi, E. Hazan, and Y. Singer. Adaptive subgradient methods for
online learning and stochastic optimization. JMLR, 2010.
[17] H. Gao, Y. Chen, K. Lee, D. Palsetia, et al. Towards Online Spam
Filtering in Social Networks. In NDSS, 2012.
[18] X. Glorot and Y. Bengio. Understanding the difficulty of training deep
feedforward neural networks. In AISTATS, 2010.
[19] Imagga. Tagging Service, 2016. https://imagga.com/.
[20] N. Jindal and B. Liu. Review spam detection. In WWW, 2007.
[21] D. Kingma and M. Welling. Auto-Encoding Variational Bayes. In ICLR,
2014.
[22] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 2015.
[23] L. Maaløe, C. K. Sønderby, S. K. Sønderby, and O. Winther. Auxiliary
Deep Generative Models. In ICML, 2016.
[24] A. Martın, A. Ashish, B. Paul, B. Eugene, C. Zhifeng, C. Craig, C. G.
S, D. Andy, D. Jeffrey, et al. Tensorflow: Large-scale machine learning
on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467,
2016.
[25] G. McLachlan and D. Peel. Finite Mixture Models. 2000.
[27] A. Peysakhovich. Reducing clickbait in facebook feed, 2016.
http://newsroom.fb.com/news/2016/08/news-feed-fyi-further-reducing-
clickbait-in-feed.
[28] Piperjaffray. Survey, 2016. http://archive.is/AA34y.
[29] E. A. Platanios and S. P. Chatzis. Gaussian Process-Mixture Conditional
Heteroscedasticity. TPAMI, 2014.
[30] I. Porteous, A. Asuncion, and M. Welling. Bayesian Matrix Factorization
with Side Information. In AAAI, 2010.
[31] M. Potthast, S. Köpsel, B. Stein, and M. Hagen. Clickbait Detection.
In ECIR, 2016.
[32] M. M. U. Rony, N. Hassan, and M. Yousuf. Diving Deep into Clickbaits:
Who Use Them to What Extents in Which Topics with What Effects?
arXiv:1703.09400, 2017.
[33] SavedYouAClick. https://twitter.com/savedyouaclick, 2014.
[34] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and R. R.
Salakhutdinov. Dropout: A Simple Way to Prevent Neural Networks
from Overfitting. JMLR, 2014.
[35] StopClickBait. http://stopclickbait.com/, 2016.
[36] G. Stringhini, M. Egele, A. Zarras, T. Holz, C. Kruegel, and G. Vigna.
B@bel: Leveraging Email Delivery for Spam Mitigation. In USENIX
Security, 2012.
[37] G. Stringhini, T. Holz, B. Stone-Gross, C. Kruegel, and G. Vigna.
BOTMAGNIFIER: Locating Spambots on the Internet. In USENIX
Security, 2011.
[38] G. Stringhini, C. Kruegel, and G. Vigna. Detecting spammers on social
networks. In CSA, 2010.
[39] T. Tieleman and G. Hinton. Lecture 6.5-RMSprop: Divide the gradient
by a running average of its recent magnitude. 2012.
[40] M. D. Zeiler and R. Fergus. Visualizing and Understanding Convolu-
tional Networks. In ECCV, 2014.