Spammer Detection: A Study of Spam Filter Comments on YouTube Videos

Abstract

This paper presents a methodology to find out the spam comments on YouTube videos. The purpose of this research is to find out the comments of those spam users, those who comment for their own promotional intentions or to detect those users' comments that have no relevancy with the given video.The monetization policy introduced by YouTube for its user's channel and advertisement of different ads on YouTube videos has attracted a large number of users. This increase in a large number of users has also lead to an increase in malicious users whose job is to create automated bots for commenting and subscription to different YouTube channels. These malicious users'comments hurt the channel publicity and also affect the normal user's experience. YouTube is also working on this issue by using different methods to limit these kinds of automated bots malicious comments by blocking those comments. These kinds of methods are ineffective so far as spammers have found out different methods to bypass those heuristic approaches. Different machine learning approaches provide somehow better classification accuracy with the introduction of new approaches to solve it better than that. In this work, different techniques used forclassification of spam comments with those of normal user comments to improve the classification and recent trend going on in this area are briefly analyzed to tackle this major issue.

Full text

Spammer Detection: A Study of Spam Filter Comments on
YouTube Videos
Rafaqat Alam Khan
Lahore Garrison University
rafaqatalam@lgu.edu.pk
Abstract:
This paper presents a methodology to find out the spam comments on YouTube videos. The purpose of this research
is to find out the comments of those spam users, those who comment for their own promotional intentions or to
detect those users’ comments that have no relevancy with the given video.The monetization policy introduced by
YouTube for its user's channel and advertisement of different ads on YouTube videos has attracted a large number
of users. This increase in a large number of users has also lead to an increase in malicious users whose job is to
create automated bots for commenting and subscription to different YouTube channels. These malicious
users’comments hurt the channel publicity and also affect the normal user’s experience. YouTube is also working on
this issue by using different methods to limit these kinds of automated bots malicious comments by blocking those
comments. These kinds of methods are ineffective so far as spammers have found out different methods to bypass
those heuristic approaches. Different machine learning approaches provide somehow better classification accuracy
with the introduction of new approaches to solve it better than that. In this work, different techniques used
forclassification of spam comments with those of normal user comments to improve the classification and recent
trend going on in this area are briefly analyzed to tackle this major issue.
Keywords:Spam, YouTube, Classification, Comment, Social Media
1 Introduction:
YouTube a video sharing website was started back
since 2005. In 2006, Google has bought YouTube
and nowadays it on Google Subsidiaries. After it
came under the umbrella of Google YouTube growth
as a YouTube, video sharing has increased
significantly. The users using YouTube can create
their own Gmail login and through this Gmail login,
they can create their own channel. Once the channel
is created, YouTube allows users to publish his own
video, rating, Comments, likes or dislikes, reporting
and subscribing to your favorite channels. According
to the recent statistics of YouTube, it has achieved
the marked of 1 billion users log in. The global
research [1] statistics of YouTube claims that around
1.9 billion users visit within a month, watching
billions of hours per day and in turn generating
billion of views per day.The 70 % of these watch
hours come from mobile devices. Around $ 2 billion
dollars, YouTube has paid over the last five years to
its YouTube channel owners.
Out of different YouTube features for its channel
users, YouTube commenting feature is one of the
important features in which users can able to
comment on other channels uploaded videos. This
powerful feature of YouTube allows the interaction
of YouTube channel owner with other users.
Introducing of such feature has also allowed other
malicious users to promote their promotional content
to know as a spam comment. These spam comments
are usually irrelevant and are generated by mostly
automated bots. The capability of bots to perform
spam comments on YouTube videos has been
discussed in [2].

Fig.1. Example of Spam Comments on YouTube Videos
A recent report [3] regarding spam comments has
been published in BBC, in which YouTube claims
that during the period of July-September 2018 they
have almost deleted 224 million malicious comments
out of billion of comments published in this 3 months
period. They have also accepted that this a really hard
work to tackle this issue and they are investing in this
area to get rid of that malicious content.
The spamming can be basically divided into two
categories i.e. link-based spam and channel-based
promotion spamming. Link-based spamming is that
type of spamming in which users comments different
links comment on published YouTube videos, on
which other users click than it redirected to malicious
webpage's. While on the other hand, channel-based
spamming is used for promotion of user channels to
get subscribers, views, and likes for his channel as
well as for his channel videos.
YouTube generally use comment blocking [4] on the
basis of HTTP URL, but this form of filtering has to
lead the spammer a new approach to publish their
links by insertion of white spaces in their link and
instead of .com they typed dot com in the form of
text to publish their link on the given video. These
problems have to lead to the machine learning
approaches to resolve such issues. The contribution
in this research is finding out of better spam
classification using different machine learning
approaches with high F1 score and better
classification accuracy.
2 Related Work
Chowdhury Rashid et al. [5] generated a lift chart by
using three different data mining models. This lift
chart finds out the lift score when compared to a
random guess. The predicted probability for the three
different data mining models i.e. Naïve Bayes,
Decision Tree and Clustering is calculated. From the
result, they have concluded that for most cases Naïve
Bayes and decision tree performed better than that of
clustering.
Tulio C. Alberto et al. [6] used different classification
algorithms i.e. Naïve Bayes, Decision tree, SVM,
Random forest and logistic regression on five

different datasets. They have achieved a confidence
level of almost 99 % on all these classifiers. Based on
these classification results they have developed their
own online tool known as TubeSpam that
automatically detects the spam message on the fly.
SaumyaGoyal et al. [7] spam message detection on
real twitter social media dataset is applied using
KNN and decision tree. The WEKA tool is used and
the metrics used for classification are precision-
recall, F measure and class, FP rate and TP rate.
SimranKanodia et al. [8] suggested a Markov
Decision process for YouTube spam message
detection and the result is compared to other data
mining tools used in this field. The Markov Decision
accuracy is 78.82 % which is quite better than those
of other data mining algorithms out of which the
maximum accuracy is obtained through the random
forest which is 72.52 %.
Abdullah O. Abdullah et al [9] WEKA tool and
python code is used for the employment of different
classification algorithms on YouTube dataset that
was generated using YouTube API. All the different
9 algorithms used have almost 90 % and above than
that accuracy. Out of these different 9 algorithms,
accuracy Adaptive Genetic Algorithm has performed
quite well and achieved an accuracy of 99.1 %.
ShreyasAiyar et al. [10] in this they have used
different machine learning algorithms along with
custom approaches i.e. N-Grams. For automated
detection of spam messages on YouTube videos they
have suggested that the character gram approach
performed better result as compared to word gram for
obtaining better classification accuracy.
3 Outline of Spam Detection
S.No
Research
Article
Techniques
Used
Available Dataset
Result
s
1
N
-
Gram
Assisted
YouTube
Spam
Comment
Detection
RF
,
SVM
,
Naïve
Bayes,N-Gram
13000 Comments
https://developers.google.com/youtube/v3/docs/c
ommentThreads#Retrieve_comments [11]
N
-
Gram
Outclassed Other
Machine Learning
Algorithms
2
A
Comparative
Analysis of
Common
YouTube
Comment
Spam Filtering
Techniques
AGA,
ICA
-
Amuse, ELM-
AE, ANN-BP,
SVM-K, K-NN,
LR, NBC, DT
100 Channels having 10,000 Samples
https://developers.google.com/youtube/v3/ [12]
Adaptive Genetic
Algorithm
Performed better
than other 8
Algorithms
3
A
Novel
Approach for
YouTube
Video Spam
Detection
using Markov
Decision
Process
Markov
Decision
Process,
Decision Tree,
Naïve Bayes,
KNN, Random
Forest, Ripper,
Clustering
50 Videos
, 2054 Instances out of which 824
was
Spam Comment and 1230 normal
Markov Decision
Process accuracy
78.52 % better
than the best RF
which was 72.82
%

4
Spam
Detection
Using KNN
and Decision
Tree
Mechanism in
Social
Network
KNN, Decision
Tree
FED Real
Dataset
http://mashable.com/2012/12/18/twitter-200-
million-active-users/ Accessed July 22, 2016 [13]
Precision Call, F
Measure and
Class, FP and TP
rate.
5
TubeSpam:
Comment
Spam Filtering
on YouTube
Naïve Bayes,
Decision Tree,
Logistic
Regression
5 Different YouTube Video Datasets
Datasets / YouTube ID / # Spam/ # Ham / Total
1) Psy9bZkp7q19f0 175 175 350
2] KatyPerryCevxZvSJLk8 175 175 350
3) LMFAO KQ6zr6kCPj8 236 202 438
4) Eminemuel Hwf8o7 U 245 203 448
5) Shakira pRpeEdMmmQ0 174 196 370
http://dcomp.sor.ufscar.br/talmeida/youtubespam
collection/ [14]
A confidence level
of 99 % on all
algorithms,
Suggested their
own App. Tube
Spam.
6
A
Data Mining
Based Spam
Detection
System For
YouTube
Naïve Bayes,
Decision Tree,
Clustering
Self Generated Data using
Tube Kit
http://www.tubekit.org/ [15]
1. No of Videos 1719
2. No of distinct users 1428
3. No of comments 10102865
4. No of ratings count 23013568
5. No of different categories 16
Lift Score is
Calculated, Naive
Bayes and
Decision Tree
performed better
having an
accuracy of 80.20
% and 82.11 %
respectively
4 YouTube Data Generation Technique
This section discussed how YouTube comments data
is generated from YouTube videos. The toolkit
TubeKit [15] is used for customized YouTube
Crawlers. This toolkit collects a variety of data from
YouTube videos. The working design of the crawler
is shown in the below fig 2.

Fig.2. Scheme for Query Based YouTube Crawling [16]
5 Conclusion and Future Work
YouTube a social networking feature website
providing one of the largest video content
publication. Recently a vast number of increases in
his popularity and its new monetization policy for
new channels have increased the number of users
with low-quality information which is known as
spam, which is posted solely for their channel
promotion or to post the malicious video link.
The automatic moderation of comment is still an
unexplored field and reason is the lack of automatic
tool detection for spam filter messages. Due to which
popular channelson YouTube try to avoid
communication with their fans on this platform and
use another platform for communication purpose.
The main goal behind this research was to find out
the different techniques and strategies to find out the
undesired comments i.e. spam messages and to
describe different dataset available for the user
working in this area. The results obtained through
this research i.e. about techniques and available
datasets description would further enhance the future
comparison.
Since for future work I would suggest that as there is
not a single method that performs well on all the
available dataset, so cascading of different machine
learning classifiers along with best feature selection
algorithms can be used to further enhance the result.
Also preprocessing of the comment text can be done
using natural language processing for text
normalization. The reason behind text normalization
is that the words used in comments are usually slang,
idioms, emoticons, symbols, and abbreviations.

6 References
[1] https://www.youtube.com/intl/en-
GB/yt/about/press/
[2] O'Callaghan, Derek et al. “Identifying
Discriminating Network Motifs in YouTube
Spam.” CoRR abs/1202.5216 (2012): n. pag.
[3] https://www.bbc.com/news/newsbeat-
46559772
[4] https://support.google.com/youtube
/answer/111870?hl=en/
[5] Chowdury, Rashid &Monsur Adnan,
Md.Nuruddin& Mahmud, G.A.N. &
Rahman, Mohammad. (2013). A data
mining based spam detection system for
YouTube. 373-378.
[6] C. Alberto, Tulio&Lochter, Johannes &
Almeida, Tiago. (2015). “TubeSpam:
Comment Spam Filtering on YouTube.”
138-143. 10.1109/ICMLA.2015.37.
[7] Goyal, Saumya& K. Chauhan, R &Parveen,
Shabnam.(2016)“Spam detection using
KNN and decision tree mechanism in social
network”522-
526.10.1109/PDGC.2016.7913250.
[8] S. Kanodia, R. Sasheendran and V. Pathari,
"A Novel Approach for Youtube Video
Spam Detection using Markov Decision
Process," 2018 International Conference on
Advances in Computing, Communications
and Informatics (ICACCI), Bangalore,
India, 2018, pp. 60-66. doi:
10.1109/ICACCI.2018.8554405.
[9] A. O. Abdullah, M. A. Ali, M. Karabatak
and A. Sengur, "A comparative analysis of
common YouTube comment spam filtering
techniques," 2018 6th International
Symposium on Digital Forensic and
Security (ISDFS), Antalya, 2018, pp. 1-5.
doi: 10.1109/ISDFS.2018.8355315
[10] ShreyasAiyar, Nisha P Shetty, “N-Gram
Assisted Youtube Spam Comment
Detection”, Procedia Computer Science,
Volume 132,2018, Pages 174-182, ISSN
1877-0509,
https://doi.org/10.1016/j.procs.2018.05.181.
[11] Youtube,YoutubeDataAPI(v3),
https://developers.google.com/youtube/v3/d
ocs/commentThreads#Retrieve_comments
[12] Data API. Available at:
https://developers.google.com/youtube/v3/
(Accessed on 22 November 2017)
[13] S. Fiegerman, “Twitter now has more than
200 million monthly active users,”
Mashable. [Online]. Available.
http://mashable.com/2012/12/18/twitter-
200-million-active-users/. Accessed July 22,
2016
[14] http://www.dt.fee.unicamp.br/~tiago//
youtubespamcollection/
[15] TubeKit, http://www.tubekit.org/. , last
accessed 24th August 2013.
[16] C. Shah, “Supporting Research Data
Collection from YouTube with TubeKit”.
Proceedings of YouTube and 2008 Election
Cycle in the United States, Amherst, MA:
April 16-17, 2009.