Suicidal Profiles Detection in Twitter
Atika Mbarek
1,2
, Salma Jamoussi
1,2
, Anis Charfi
3
and Abdelmajid Ben Hamadou
1,2
1
Multimedia InfoRmation Systems and Advanced Computing Laboratory (MIRACL), University of Sfax, Tunisia
2
Digital Research Center of Sfax DRCS, 3021, Sfax, Tunisia
3
Carnegie Mellon University in Qatar, Doha, Qatar
Abdelmajid.benhamadou@isimsf.rnu.tn
Keywords: Suicide, Twitter, User Profile, Machine Learning.
Abstract: About 800 000 people commit suicide every year and detecting suicidal people remains a challenging issue
as mentioned in a number of suicide studies. With the increased use of social media, we witnessed that
people talk about their suicide plans or attempts in public on these networks. This paper addresses the
problem of suicide prevention by detecting suicidal profiles in social networks and specifically twitter. First,
we analyse profiles from twitter and extract various features including account features that are related to
the profile and features that are related to the tweets. Second, we introduce our method based on machine
learning algorithms to detect suicidal profiles using Twitter data. Then, we use a profile data set consisting
of people who have already committed suicide. Experimental results verify the effectiveness of our
approachin terms of recall and precision to detect suicidal profiles. Finally, we present a Java based
prototype of our work that shows the detection of suicidal profiles.
1 INTRODUCTION
Social media has changed the world. It has become an
everyday part of our lives. Many people are nowadays
active on several popular social networks such as
Facebook, twitter, Instagram, etc. They share photos
and posts on their daily life and experiences such as
their food, their clothes, and their trips. Some people
are more active on social networks, while others are
less so.
On the other hand, social networks can reflect
different social phenomena such as diseases,
depression, suicide, etc. In particular, suicide is a
complex and dangerous phenomenon that should be
considered and studied in order to reduce mortality
rates. A recent study
1
revealed that close to 800 000
people commit suicide every year, which means one
person every 40 seconds. Thus, this growing
phenomenon presents one of the biggest challenges
the world is facing today. Understanding the
symptoms related to suicidal tendencies is important
to prevent such deaths.
1
https://www.who.int/mental_health/prevention/suicide/sui
cideprevent/en/
In this respect, many studies on suicide prevention
have become more prevalent in recent years. Indeed,
one of the greatest things that characterize social
networks is their use in extracting emotional thoughts
and feelings of depression. For that reason, many
researchers rely on social networks to study suicide.
As an example, Twitter has become a very popular
social network where millions of users share their
opinions and feelings using short texts called tweets,
which contains semantic expressions such as
emoticons, hashtags, special characters, etc.
Consequently, twitter provides a rich source of data
for text mining.
Most suicidal people who are active in social
networks give signals of their intentions. For example,
they make statements such as "I want to kill myself,"
"I hate my life", "I have lived long enough "or “I’m so
tired”. The best way to prevent their suicide is to
catch these signals and predict other hidden signals
behind their posting content in order to react to them
and take appropriate actions.
Generally, a user in twitter is characterized by a
profile and a set of tweets. The profile features
describe his/her persona such as name, age, location,
date of birth. On the other hand, tweets refer to the
content shared by the user such as text, photos or
videos. Some existing works (Jain et al., 2013) in this
Mbarek, A., Jamoussi, S., Charfi, A. and Ben Hamadou, A.
Suicidal Profiles Detection in Twitter.
DOI: 10.5220/0008167602890296
In Proceedings of the 15th International Conference on Web Information Systems and Technologies (WEBIST 2019), pages 289-296
ISBN: 978-989-758-386-5
Copyright
c
2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
289
context utilize publicly shared attributes including
name, gender, location, and other information to
identify user profiles in social networks. However,
due to the privacy settings, user’s attributes are not
available in many cases and this makes these existing
works fragile. In addition, some researchers address
the problem of suicide only through tweets (Kavuluru
et al., 2016), (Colombo et al., 2015). However, even
though tweets contain rich information that can
identify users, they can miss some significant details
that maybe available on the user profile public
attributes and may contribute to a higher accuracy of
suicide detection. Apart from these works, we utilize
in our approach both user shared information that we
call account features and tweets as an attempt to solve
the problem of suicidal profiles detection. First,
posted tweets pose important challenges to infer more
information about users. The most relevant challenge
is semantic features that are difficult to extract
directly from user’s posted tweets such as stylometry,
writing style, sentiments, emojis, hashtags, n-grams,
etc. Instead of many existing studies that ignore these
features to identify users, we analyse tweets and
extract as much as possible of semantic features.
Second, adding account features to the user’s posted
tweets can help to improve the suicide detection task
since they may reflect the habits and characteristics of
users.
Although there are many studies (Sueki, 2014),
(O'dea, 2018) that focused on the particular problem
of suicidality detection in social networks, they do not
take into account the profile itself. They only
considered suicide related-communication with the
aim of classifying text relating to suicide. However,
the biggest challenge for the suicide task is how to
detect users who want to commit suicide from their
public profiles in social networks.
In this paper, we consider the challenge of suicidal
profiles detection in Twitter. We analyse posted
tweets to extract semantic features including
linguistic, emotional, stylometric, etc. These features
allow us to distinguish between the writing styles of
different users and thus to facilitate the final
classification of users into suicidal or not suicidal. In
particular, posted tweets contain temporal information
that can indicate the real time of user’s posting. Such
information is very relevant to enrich the user
identification and improve the suicide detection. We
also use account features related to publicly shared
information such as profile photo, location,
biography, followees, etc. We exploit these features to
infer other implicit ones and build a rich profile that
can help us to predict suicidal users. We adopt
different data mining tools and techniques for the
extraction process. We also introduce a supervised
machine learning model to learn the features
identifying each user. Moreover, we adopt several
classification techniques to classify profiles into
suicidal and not suicidal. We apply our method to a
data set collected from Twitter and including profiles
whose owners committed suicide.
The rest of this paper is organized as follows:
Section 2 discusses related work. Our method of
suicidal profiles detection is explained in Section 3,
which also presents the collection of data from twitter
based on tweets and account features. Section
4reports on evaluation. Section 5 concludes the paper
and outlines directions for future work.
2 RELATED WORK
Social media have become increasingly popular and
the number of active users continues to increase.
Several phenomena such as suicide are now visible on
social media. To address suicide and reduce the
related mortality rates, many studies were conducted
on suicidality in social networks.
Kavuluru et al., 2016 conducted a suicide study
by classifying text relating to suicide on Twitter. They
built a set of account classifiers using lexical,
structural, emotive and psychological features
extracted from Twitter posts. Their aim was to
distinguish between the more worrying content, such
as suicidal ideation, and other suicide-related topics.
Other studies (Kavuluru et al., 2014) have focused
on the writing styleusing the LIWC tool as a sampling
technique to identify ’sad’ Twitter posts that were
subsequently classified using a machine learning
classifier into levels of distress on an ordinal scale,
with around 64% accuracy in the best-case.
Additionally, (Birjali et al., 2017) based their work on
WordNet to analyse semantically Twitter data. They
address the lack of terminological resources related to
suicide by constructing a vocabulary associated with
suicide.
A case study (O'dea et al., 2015) used both human
coders and a machine classifier to confirm that
Twitter is used by individuals to express suicidality
and that it is possible to distinguish the level of
concern among suicide-related tweets.
In another work, (De Choudhury et al., 2016)
considered online platforms such as Reddit and
applied topic analysis and linguistic features to
identify behavioural shifts and mental health issues
such as suicidal ideation, thus highlighting the risks of
supposedly helpful messages in such online forums.
Furthermore, (Colombo et al., 2015) investigated the
WEBIST 2019 - 15th International Conference on Web Information Systems and Technologies
290
characteristics of the authors of Tweets containing
suicidal thinking, through the analysis of their online
social network relationships rather than focusing on
the text of their posts.
More recently, (O'dea et al., 2018) used a dataset
of suicide related posts to study how Twitter users
respond to suicide-related content compared to non-
suicide related content. They found that the rate of
reply to the suicide-related posts was significantly
faster than that one for non-suicide related posts, with
the average reply occurring within 1 hour. Finally,
(Braithwaite et al., 2016) classified text from Twitter
users as suicidal or non-suicidal using affective
markers and machine classification algorithms –
stopping short of examining texts for other forms of
suicidal communication.
Existing works on suicide prevention are mainly
focused on identifying suicidal thinking or ideation
and detecting suicidal posts. However, there are no
significant research works that focused in particular
on suicidal profile detection. Thus, our study aims to
contribute to the literature on understanding
communication on the topic of suicide in social
networks by detecting suicidal profiles on Twitter.
3 SUICIDAL PROFILE
DETECTION
In this work, we propose a method for detecting
suicidal profiles. First, we analyse a number of
profiles from the social network Twitter through
exploiting the maximum of available data. Then, we
adopt several features to distinguish between suicidal
and not suicidal profiles. These features can be
explicitly extracted from the user profile or implicitly
inferred using different data mining tools and
techniques. Here, we focus on emotional features and
sentiment analysis, which gives indications about the
psychological state of suicidal profiles. We further
use account features to identify users through the
shared information on their profiles. We finally
present each user as a vector that integrates all the
used features.
3.1 Data Preparation
Before proceeding to the analysis of the profiles and
the extraction of features, it seemed necessary to us to
clean and normalize the collected data. Generally,
posted tweets are short and noisy. For instance, the
language used is very informal, with Unicode
characters, punctuation, poor spelling, acronyms,
URLs, and abbreviations. Thus, to make the user’s
content look clearer and to improve the text analysis,
we made a dictionary for the useless stop-words and
created an R code that eliminates all noisy words from
the original text.
3.2 Features Extraction
Clearly, gathering rich information about users is
crucial for providing a high quality of suicide risk
detection. Moreover, several types of features can
lead to more accurate suicidal detection. This allows
us to decide which profile can be suicidal. Therefore,
extracting features from twitter profiles is a necessary
step for the classification process. To do so, we
employed some data mining tools and techniques in
order to infer implicit information that were not given
explicitly by the user. We further used the Linguistic
Inquiry and Word Count LIWC text analysis software
(Pennebaker, 2001), to extract more relevant features
associated with emotions. The great advantage of this
tool is that it analyses text files on a word-by-word
basis using an internal dictionary and computes the
percentage of words in a text that are in each of these
linguistic or psychological categories. Thus, it helped
us to enrich the user’s information especially on the
emotional side, which is very important for the
suicide topic. We considered in our work two types of
features: account features and features based on
tweets.
3.2.1 Account Features
In this type of features, we only consider information
related to the profile. In other words, we do not
consider tweets to identify users. Indeed, we use
information that can be explicitly extracted from the
profile. We consider three categories of account
features according to their consistency.
Explicit Features.
Explicit features are those publicly shared by the user
in his/her profile.
Table 1: Explicit features availability.
Twitter feature
Availability
Language
80%
Country
66%
Created profile
100%
Friend’s number
100%
Profile description
90%
Profile photo
100%
Suicidal Profiles Detection in Twitter
291
They refer to the user’s name, language, country,
profile creation date, number of friends, profile
description and profile photo. Some features are
almost available while some others are sometimes
missing.
Table 1 describes the availability of explicit attributes
in the social network Twitter.
Facial Features.
The main limitation of twitter is the fact that some
user’s attributes are missing. In particular, age and
gender, which are relevant information for identifying
users, are not available in twitter. In order to deal with
this issue, we used the picture profile to extract facial
features such as gender and age. We used Microsoft
Face API
2
, a cloud-based service that provides the
most advanced face algorithms (MAHESHWARI,
2017) to extract through a photo various facial
attributes including gender, age, smile, facial Hair
beard, facial Hair moustache, and facial Hair
sideburns.
Followees.
Finding posts and profiles which the user follows may
be a relevant indication to know which kind of
profiles the user is interested in and to know about
his/her interactions with other profiles. Consequently,
if the user has suicidal ideations, normally she follows
topics related to suicide. In addition, this information,
allows us to know the degree of sociability of the user
with people in the social network twitter. Thus, we
collected posts with which the user interact with
through likes, comments or retweets, and profiles who
share these posts.
Table 2: Information extracted from followees.
Followees
Age ranges
Histogram of photos
Topics of Interest
Sentiment analysis
Through the user’s followees, we extract other
information that describes more the profiles followed
by the user to know if they really have an influence on
his/her thoughts. For example, the histogram of
photos allows us to know whether they use images
with light colours or dark images. In particular,
depressed people or those who have suicidal thoughts
prefer to put dark images in their profiles. In addition,
2
https://docs.microsoft.com/en-us/azure/cognitive-
services/face/quickstarts/csharp
we extract the sentiment analysis and topics of
interest from the tweets posted by the user’s followees
to identify the emotional state of these profiles. Also,
the age ranges of the user’s followees present relevant
information. In other words, people with the same age
or close in age have similar interests as a result of
age-related life events.
3.2.2 Features based on Tweets
Linguistic Features.
Linguistic features are very important to distinguish
between written styles of users. Indeed, many studies
in suicidal ideation (Sueki, 2014), (BURNAP, 2017)
focused on linguistic features to prevent suicide. It
seems obvious that each user has a unique writing
style. In social networks, some users may use similar
writing style. For example, users that post emotional
text generally use some particular characteristics such
as elongation, adjectives, exclamations, etc.
Therefore, extracting these features may highly
increase the chances of detecting suicidal profiles. In
this context, there are various writing styles that can
be extracted from Twitter, and that contain different
characteristics.
Table 3: Writing style description.
Writing style features
Description
LIWC features
e.g. adjectives, pronouns, adverbs,
health, death, etc.
Special characters
Percentage of used special characters
(e.g. $, %,&, (, ), *, +, _, /, <, =, >, @,
etc)
Frequent words
Number of frequent words repeated
more than 5 times
N-grams
Number of frequent n-grams repeated
more than 3 times (e.g. No More , want
to kill, terrible times, don’t want,etc)
Elongations
Percentage of used elongations(e.g.
nooooo, ohhhhhhh, !!!!!!!!!, loooool,
etc)
Sentences length
Average length of used sentences
Words length
Average length of used words
Writing language
Number of used languages
Htags
Percentage of used htags
We employed the LIWC tool to extract various
linguistic features as described in table 3. For the
other features, we implemented an R code that
computes their values using the collected tweets.
Emotional Features.
Most suicidal users suffer from mental health
problems due to psychiatric disorder, social problems,
substance abuse, etc. Therefore, their posted tweets
are generally associated with depression terms such as
WEBIST 2019 - 15th International Conference on Web Information Systems and Technologies
292
death, mad, killing, lonely, etc. However, the frequent
occurrence of these terms in their posted texts may
increase the risk of suicidal ideation. Therefore,
extracting this type of features has helped so far to
reveal suicidal users. In our work, we focused on two
main emotional features: sentiment analysis and
emojis.
Sentiment Analysis.
Sentiment analysis works best on text that has a
subjective context such as suicide. In this context,
many studies (Pestian, 2012), (BIRJALI, 2017) used
sentiment analysis to target the problem of suicide
since emotions are closely related to sentiment.
Typically, they classify feelings according to their
polarities: either positive, negative or neutral. Having
the same aim, we adopt Open NLP
3
as a machine
learning based toolkit for processing natural language,
which understands the language used in a text and
uncovers the sentiment behind it. We also added two
significant attributes associated with sentiment named
positive terms and negatives ones. To that end, we
prepared a rich dictionary that contains positive and
negative terms collected from differentsites. We also
added a dictionary vocabulary of words called
opinion lexicon
4
, which includes around 6800
negative and positive English words. The sentiment
analysis features are shown in Table4.
Table 4: Sentiment analysis features description.
Sentiment analysis
features
Description
Positive sentiment
Percentage of positive used sentences
Negative
sentiment
Percentage of negative used sentences
Neutral sentiment
Percentage of neutral used sentences
Positive terms
Number of positive used terms (e.g.
happy, enjoy, well, wonderful, well, etc)
Negative terms
Number of negative terms (e.g. sad,
suffer, depression, mad, etc)
Emojis.
Emojis or emoticons are often used to show various types of
emotions. However, using typical emojis such as the happy
and sad emojis may not give detailed emotional state of
users. Clearly, the more we specify the type of emojis the
more we get precise recognition of emotions. Based
on(Parrott, 2001), we employ six primary emotions
including love, joy, surprise, anger, sadness, and fear.
Emojis features are described in Table 5.
3
https://cran.rproject.org/web/packages/openNLP/index.ht
ml
4
https://www.cs.uic.edu/~liub/FBS/sentiment-
analysis.html
Table 5: Emojis features description.
Examples
:-* :* <3 ♥ � � �
xD :-) :) :D :o) :] :3 :c) :> =]
8) ;-) :-P XP
:-O :O :-o : o :-0 8-0
:-J >:( >:O
:-( :( :'( :'-(
%-) %) v.v
Some emojis include simple characters that are
typically used in text while others include Unicode
characters that are difficult to explicitly extract. To
deal with that, we exploit recent emoticons and smiley
faces existing in some sentiment sites with unicode.
Temporal Features.
Collecting data from Twitter provides us temporal
information that indicates the real time of a user's
posting.
Table 6: Timeline features description.
Description
Percentage of posting tweets 20:00
to 04:00
Percentage of posting tweets 04:01
to 12:00
Percentage of posting tweets 12:01
to 20:00
Percentage of posting tweets all the
week-day except Saturday and
Sunday
Percentage of posting tweets on
Saturday and Sunday
Percentage of posting in winter
Percentage of posting in summer
Percentage of posting in spring
Percentage of posting in autumn
Percentage of posting per day
Percentage of posting per week
Percentage of posting per month
According to (Xiangnan, 2013), a user usually posts
his/her content on different social networks at similar
time slots. Such temporal information is very relevant
for user identification. For example, some users like
to post their content at night, while other users share
their posts in the morning. Also, there are users who
are very active on the weekend unlike other users who
share posts frequently on ordinary days. Thus, for
each user, we can extract the exact time of his/her
publicly posted tweets. Through this information, we
collect 12 types of temporal features as described in
Table 6.
Suicidal Profiles Detection in Twitter
293
4 EVALUATION
4.1 Dataset
We spent 2months to collect twitter profiles whose
owners committed suicide. We referred to the
TWEET HEREAFTER
5
site, which contains some
users that uttered a final word on their profiles a few
time prior their suicide. We also searched for
popular persons that committed suicide worldwide
and checked if they had twitter profiles. Our final
dataset consisted of 115 suicidal profiles and 172 not
suicidal profiles. Statistics on this set are provided in
Table7.
Table 7: Statistical information of dataset.
Twitter
Suicidal profiles
Non suicidal profiles
Profile numbers
115
172
Male numbers
41
81
Female numbers
74
91
Average number of
tweets per user
84
102
4.2 Learning Process
In order to demonstrate the effectiveness of our
work in detecting suicidal profiles, we need to
compare it with prior works. However, such
comparisons are not evident because prior works did
not target the profile, but rather focused on detecting
suicidal text. On the other hand, feature descriptions
are different and not all works use the same features.
Therefore, we made a comparison of methods using
another way. We conducted two types of
experiments to show the effectiveness of our work.
First, we used only features based on tweets to
detect suicidal profiles. Then, we added account
features and presented the difference between the
results.
We used the collected profiles in order to train
and test a number of machine classifiers to classify
profiles into suicidal and not suicidal. We adopted a
supervised machine learning approach based on
various features as described in Section 3. We used
Weka
6
as a data mining tool to extract all useful
information for the classification of suicidal profiles
according to the machine learning algorithms
5
http://thetweethereafter.com/?page=2&s=death_desc&fbc
lid=IwAR34cC1kZFiCzt0Mapn_lX4MuAGdhyVGNfi
nkZRlHJE5VVZfompbCXz3sD4
6
https://www.cs.waikato.ac.nz/~ml/weka/
implemented in Weka. We selected five classifiers
including BayesNet, Adaboost, J48, SMO and
Random Forest. Experiments were carried out with
10-fold cross validation on the training data.
In order to evaluate further of our method in
identifying suicidal profiles, we implemented a web
based java application that shows if a given profile is
suicidal or not. Given a twitter screen name, the
application extracts all features related to the
targeted profile and then returns the prediction
result. Figure 1 shows the suicide prediction result of
a profile related to a user who committed suicide
recently.
Figure 1: Suicidal profile detection exemple.
4.3 Experiments and Results
4.3.1 Using Only Tweet based Features
Table 8 presents the precision, recall, and F-measure
of our model in identifying suicidal profiles when
we use only features based on tweets. The
experiments were run with a 10-fold cross
validation. As shown in table 8, the best reached F-
Measure is 77% with the random forest classifier.
The precision of the SMO classifier with a
PolyKernel function was74%.
Table 8: Classification results using features based on
tweets.
Classifiers
Precision
Recall
F-measure
Bayes Net
70%
70%
70%
Adaboost
71%
71%
71%
SMO
74%
74%
74%
J48
70%
70%
70%
Random
Forest
77%
77%
77%
4.3.2 Using All Features
Then, we tested our method with all the features
including account features and features based on
WEBIST 2019 - 15th International Conference on Web Information Systems and Technologies
294
tweets. The experiments were run with a 10-fold
cross validation. Table 9 shows the obtained results.
Table 9: Classification results using all features.
Classifiers
Precision
Recall
F-mesure
Bayes Net
73%
74%
73%
Adaboost
78%
78%
78%
SMO
79%
79%
78%
J48
80%
81%
81%
Random
Forest
83%
83%
83%
The results show that the best performing
classifier in terms of precision is Random Forest
yielding a value of 83%. When using all features, the
results improved with all classifiers. In particular,
J48 classifier increased to 80% in terms of precision
instead of 70% when using only tweet based
features.
Furthermore, we test our method with a testing
set including 10 suicidal profiles and 10 not suicidal
ones as shown in Table 10.
Table 10: Classification results using testing test.
Classifiers
Precision
Recall
F-measure
Bayes Net
72%
80%
76%
Adaboost
81%
90%
85%
SMO
72%
80%
76%
J48
72%
80%
76%
Random
Forest
81%
90%
85%
The results show that our method performed well
on the profiles of the testing set. As shown in Table
10, we were able to a reach a recall of 90% with the
classifiers Random Forest and Adaboost, which
means that we could correctly classify 9 profiles out
of 10. This proves the effectiveness of our method to
identify suicidal profiles.
5 CONCLUSIONS
In this paper, we worked on detecting user profiles
that are at risk of suicide. We worked on twitter and
defined a detection model using a set of rich features
including linguistic, emotional, facial, timeline as
well as public features to identify twitter profiles.
We used several machine learning methods (mainly
classifiers) for the suicidal detection. Moreover, we
implemented a Java based tool to detect suicidal
profiles. To evaluate our work, we conducted a
series of experiments using a data set of profiles that
committed suicide. Results were promising with an
average recall of 86%.
As future work, we aim at improving the results
of the suicidal profiles detection by determining
more precisely their degree of suicidality.
Furthermore, we would like to target other social
media platforms such as Facebook and Instagram.
ACKNOWLEDGEMENTS
This publication was made possible by NPRP grant
#9-175-1-033 from the Qatar National Research
Fund (a member of Qatar Foundation). The
statements made herein are solely the responsibility
of the authors.
REFERENCES
H. Sueki, The association of suicide-related twitter use
with suicidal behaviour: across-sectional study of
young internet users in japan, J. Affect Disord. 170
(2014)155–160.
Burnap, Pete, et al. Multi-class machine classification of
suicide-related communication on Twitter. Online
social networks and media, 2017, 2: 32-44.
Pennebaker, James W.; FRANCIS, Martha E.; BOOTH,
Roger J. Linguistic inquiry and word count: LIWC
2001. Mahway: Lawrence Erlbaum Associates, 2001,
71.2001: 2001.
Jain, Paridhi, PonnurangamKumaraguru, and
AnupamJoshi. "@ iseek'fb.me': Identifyingusers
across multiple online social networks." Proceedings
of the 22nd international conference on World Wide
Web. ACM, 2013.
J.P. Pestian, P. Matykiewicz, M. Linn-Gust, B. South, O.
Uzuner, J. Wiebe, K.B. Cohen, J. Hurdle, C. Brew,
Sentiment analysis of suicide notes: A shared task,
Biomed. Inf. Insights 5 (Suppl 1) (2012) 3.
Birjali, Marouane; BENI-HSSANE, Abderrahim; Erritali,
Mohammed. Machine learning and semantic sentiment
analysis based algorithms for suicide sentiment
prediction in social networks. Procedia Computer
Science, 2017, 113: 65-72.
W.G. Parrott (ed.), Emotions in Social Psychology:
Essential Readings, Key Reading in Social
Psychology, Psychology Press, Philadelphia, PA 2001.
Xiangnan Kong, Jiawei Zhang, Philip S. Yu, Inferring
anchor links across multiple heterogeneous social
networks, in: Proceedings of the 22nd ACM
international conference on Information & Knowledge
Management, 2013, pp. 179–188
Birjali, Marouane; BENI-HSSANE, Abderrahim; Erritali,
Mohammed. Machine learning and semantic sentiment
analysis based algorithms for suicide sentiment
prediction in social networks. Procedia Computer
Science, 2017, 113: 65 72.
Suicidal Profiles Detection in Twitter
295
O'dea, Bridianne, et al. Detecting suicidality on
Twitter. Internet Interventions, 2015, 2.2: 183-188.
O'dea, Bridianne, et al. The rate of reply and nature of
responses to suicide-related posts on Twitter. Internet
interventions, 2018, 13: 105-107.
R. Kavuluru, M. Ramos-Morales, T. Holaday, A.G.
Williams, L. Haye, J. Cerel, Classification of helpful
comments on online suicide watch forums., in: BCB,
2016, pp. 32–40.
M. De Choudhury, E. Kiciman, M. Dredze, G.
Coppersmith, M. Kumar, Discovering shifts to suicidal
ideation from mental health content in social media,
in: Proceedings of the 2016 CHI Conference on
Human Factors in Computing Systems, ACM, 2016,
pp. 2098–2110.
G.B. Colombo, P. Burnap, A. Hodorog, J. Scourfield,
Analysing the connectivity and communication of
suicidal users on Twitter, Comput. Commun. 73
(2016) 291–300. Online Social Networks, doi:
10.1016/j.comcom.2015.07.018.
C. Homan, R. Johar, T. Liu, M. Lytle, V. Silenzio, C.
OvesdotterAlm, Towardmacro-insights for suicide
prevention: analyzing fine-grained distress at scale,in:
Proceedings of the Workshop on Computational
Linguistics and ClinicalPsychology, Association for
Computational Linguistics, Baltimore, Maryland,USA,
2014, pp. 107–117.
R.S. Braithwaite, C. Giraud-Carrier, J. West, D.M. Barnes,
L.C. Hanson, Validating machine learning algorithms
for Twitter data against established measures of
suicidality, JMIR Ment. Health 3 (2) (2016) e21,
doi:10.2196/mental.4822.
Maheshwari, Karan, et al. Facial Recognition Enabled
Smart Door Using MicrosoftFace API. arXiv preprint
arXiv:1706.00498, 2017.
WEBIST 2019 - 15th International Conference on Web Information Systems and Technologies
296