This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS 1

Suicidal Ideation Detection: A Review of Machine

Learning Methods and Applications
Shaoxiong Ji , Shirui Pan , Member, IEEE, Xue Li, Member, IEEE, Erik Cambria , Senior Member, IEEE,
Guodong Long , and Zi Huang , Member, IEEE

Abstract— Suicide is a critical issue in modern society. Early The reasons that people commit suicide are complicated.
detection and prevention of suicide attempts should be addressed People with depression are highly likely to commit suicide, but
to save people’s life. Current suicidal ideation detection (SID) many without depression can also have suicidal thoughts [2].
methods include clinical methods based on the interaction
between social workers or experts and the targeted individuals According to the American Foundation for Suicide Pre-
and machine learning techniques with feature engineering or vention (AFSP), suicide factors fall under three categories:
deep learning for automatic detection based on online social health factors, environmental factors, and historical factors [3].
contents. This article is the first survey that comprehensively Ferrari et al. [4] found that mental health issues and substance
introduces and discusses the methods from these categories. use disorders are attributed to the factors of suicide. O’Connor
Domain-specific applications of SID are reviewed according to
their data sources, i.e., questionnaires, electronic health records, and Nock [5] conducted a thorough review of the psychology
suicide notes, and online user content. Several specific tasks and of suicide and summarized psychological risks as personality
data sets are introduced and summarized to facilitate further and individual differences, cognitive factors, social factors, and
research. Finally, we summarize the limitations of current work negative life events.
and provide an outlook of further research directions. Suicidal ideation detection (SID) determines whether the
Index Terms— Deep learning, feature engineering, social con- person has suicidal ideation or thoughts by given tabular data
tent, suicidal ideation detection (SID). of a person or textual content written by a person. Due to the
advances in social media and online anonymity, an increasing
I. I NTRODUCTION number of individuals turn to interact with others on the
Internet. Online communication channels are becoming a new
M ENTAL health issues, such as anxiety and depression,
are becoming increasingly concerned in modern soci-
ety, as they turn out to be especially severe in developed
way for people to express their feelings, suffering, and suicidal
tendencies. Hence, online channels have naturally started to
countries and emerging markets. Severe mental disorders with- act as a surveillance tool for suicidal ideation, and mining
out effective treatment can turn to suicidal ideation or even social content can improve suicide prevention [6]. Strange
suicide attempts. Some online posts contain much negative social phenomena are emerging, e.g., online communities
information and generate problematic phenomena, such as reaching an agreement on self-mutilation and copycat suicide.
cyberstalking and cyberbullying. Consequences can be severe For example, a social network phenomenon called the “Blue
and risky since such lousy information is often engaged Whale Game”1 in 2016 uses many tasks (such as self-harming)
in some form of social cruelty, leading to rumors or even and leads game members to commit suicide in the end. Suicide
mental damage. Research shows that there is a link between is a critical social issue and takes thousands of lives every year.
cyberbullying and suicide [1]. Victims overexposed to too Thus, it is necessary to detect suicidality and prevent suicide
many negative messages or events may become depressed and before victims end their life. Early detection and treatment
desperate; even worse, some may commit suicide. are regarded as the most effective ways to prevent potential
suicide attempts.
Manuscript received October 17, 2019; revised April 26, 2020, June 22,
2020, and August 16, 2020; accepted August 31, 2020. (Corresponding Potential victims with suicidal ideation may express their
author: Shaoxiong Ji.) thoughts of committing suicide in fleeting thoughts, suicide
Shaoxiong Ji is with the Department of Computer Science, Aalto University, plans, and role-playing. SID is to find out these risks of
02150 Espoo, Finland, and also with the School of Information Technology
and Electrical Engineering, The University of Queensland, Brisbane, QLD intentions or behaviors before tragedy strikes. A meta-analysis
4072, Australia (e-mail: [email protected]). conducted by McHugh et al. [7] shown statistical limita-
Shirui Pan is with the Faculty of Information Technology, Monash Univer- tions of ideation as a screening tool but also pointed out
sity, Melbourne, VIC 3800, Australia (e-mail: [email protected]).
Xue Li and Zi Huang are with the School of Information Technology and that people’s expression of suicidal ideation represents their
Electrical Engineering, The University of Queensland, Brisbane, QLD 4072, psychological distress. Effective detection of early signals of
Australia (e-mail: [email protected]; [email protected]). suicidal ideation can identify people with suicidal thoughts
Erik Cambria is with the School of Computer Science and Engineering,
Nanyang Technological University, Singapore 639798 (e-mail: cambria@ and open a communication portal to let social workers mit-
ntu.edu.sg). igate their mental issues. The reasons for suicide are com-
Guodong Long is with the Faculty of Engineering and IT, University of plicated and attributed to a complex interaction of many
Technology Sydney, Ultimo, NSW 2007, Australia (e-mail: guodong.long@
uts.edu.au).
Digital Object Identifier 10.1109/TCSS.2020.3021467 1 https://thesun.co.uk/news/worldnews/3003805

2329-924X © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

Fig. 1. Categorization of suicide ideation detection: methods and domains. The left part represents method categorization, while the right part shows the
categories of domains. The arrow and solid point indicate subcategories.

factors [5], [8]. To detect suicidal ideation, many researchers is one of the potential applications for social good and should
conducted psychological and clinical studies [9] and classified be addressed to improve people’s wellbeing meaningfully.
responses of questionnaires [10]. Based on their social media The research problems include feature selection on tabular
data, artificial intelligence (AI) and machine learning tech- and text data and representation learning on natural language.
niques can predict people’s likelihood of suicide [11], which Many AI-based methods have been applied to classify suicide
can better understand people’s intentions and pave the way risks. However, there remain some challenges. There are a
for early intervention. Detection on social content focuses on limited number of benchmarks for training and evaluating
feature engineering [12], [13], sentiment analysis [14], [15], SID. AI-powered models, sometimes, learn statistical clues but
and deep learning [16]–[18]. Those methods generally require fail to understand people’s intentions. Moreover, many neural
heuristics to select features or design artificial neural net- models are lack of interpretability. This survey reviews SID
work (ANN) architectures for learning rich representation. The methods from the perspective of AI and machine learning
research trend focuses on selecting more useful features from and specific domain applications with social impact. The
people’s health records and developing neural architectures to categorization from these two perspectives is shown in Fig. 1.
understand the language with suicidal ideation better. This article provides a comprehensive review of the increas-
Mobile technologies have been studied and applied to ingly important field of SID with machine learning methods.
suicide prevention, for example, the mobile suicide interven- It proposes a summary of current research progress and an
tion application iBobbly [19] developed by the Black Dog outlook of future work. The contributions of our survey are
Institute.2 Many other suicide prevention tools integrated with summarized as follows.
social networking services have also been developed, including 1) To the best of our knowledge, this is the first survey that
Samaritans Radar3 and Woebot.4 The former was a Twitter conducts a comprehensive review of SID, its methods,
plugin that was later discontinued because of privacy issues. and its applications from a machine learning perspective.
For monitoring alarming posts, the latter is a Facebook chatbot 2) We introduce and discuss the classical content analysis
based on cognitive behavioral therapy and natural language and modern machine learning techniques, plus their
processing (NLP) techniques for relieving people’s depression application to questionnaires, EHR data, suicide notes,
and anxiety. and online social content.
Applying cutting-edge AI technologies for SID inevitably 3) We enumerate existing and less explored tasks and
comes with privacy issues [20] and ethical concerns [21]. discuss their limitations. We also summarize existing
Linthicum et al. [22] put forward three ethical issues, including data sets and provide an outlook of future research
the influence of bias on machine learning algorithms, the pre- directions in this field.
diction on time of suicide act, and ethical and legal questions The remainder of this article is organized as follows.
raised by false positive and false negative prediction. It is Methods and applications are introduced and summarized
not easy to answer ethical questions for AI as these require in Sections II and III, respectively. Section IV enumerates
algorithms to reach a balance between competing values, specific tasks and some data sets. Finally, we have a discussion
issues, and interests [20]. and propose some future directions in Section V.
AI has been applied to solve many challenging social
problems. Detection of suicidal ideation with AI techniques
II. M ETHODS AND C ATEGORIZATION
2 https://blackdoginstitute.org.au/research/digital-dog/programs/ibobbly-app
3 https://samaritans.org/about-samaritans/research-policy/internet- Suicide detection has drawn the attention of many
suicide/samaritans-radar researchers due to an increasing suicide rate in recent years
4 https://woebot.io and has been studied extensively from many perspectives.

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

JI et al.: SID: A REVIEW OF MACHINE LEARNING METHODS AND APPLICATIONS 3

The research techniques used to examine suicide also span risk [38], social media content detection, and speech patterns
many fields and methods, for example, clinical methods with analysis [39].
patient–clinic interaction [9] and automatic detection from
user-generated content (mainly text) [12], [17]. Machine learn- B. Feature Engineering
ing techniques are widely applied for automatic detection. The goal of text-based suicide classification is to deter-
Traditional suicide detection relies on clinical meth- mine whether candidates, through their posts, have suicidal
ods, including self-reports and face-to-face interviews. ideations. Machine learning methods and NLP have also been
Venek et al. [9] designed a five-item ubiquitous questionnaire applied in this field.
for the assessment of suicidal risks and applied a hierarchical 1) Tabular Features: Tabular data for SID consist of
classifier on the patients’ response to determine their suici- questionnaire responses and structured statistical informa-
dal intentions. Through face-to-face interaction, verbal and tion extracted from websites. Such structured data can be
acoustic information can be utilized. Scherer [23] investigated directly used as features for classification or regression.
the prosodic speech characteristics and voice quality in a Masuda et al. [40] applied logistic regression to classify
dyadic interview to identify suicidal and nonsuicidal juveniles. suicide and control groups based on users’ characteristics
Other clinical methods examine the resting state heart rate and social behavior variables. The authors found variables
from converted sensing signals [24] and classify the functional such as community number, local clustering coefficient, and
magnetic resonance imaging-based neural representations of homophily have a more substantial influence on suicidal
death- and life-related words [25] and event-related instigators ideation in an SNS of Japan. Chattopadhyay [41] applied
converted from EEG signals [26]. Another aspect of clinical pierce suicidal intent scale (PSIS) to assess suicide factors and
treatment is the understanding of the psychology behind conducted regression analysis. Questionnaires act as a good
suicidal behavior [5], which, however, relies heavily on the source of tabular features. Delgado-Gomez et al. [42] used
clinician’s knowledge and face-to-face interaction. Suicide risk the international personality disorder examination screening
assessment scales with clinical interview can reveal informa- questionnaire and the Holmes–Rahe social readjustment rat-
tive cues for predicting suicide [27]. Tan et al. [28] conducted ing scale (SRRS). Chattopadhyay [43] proposed to apply a
an interview and survey study in Weibo, a Twitter-like service multilayer feedforward neural network, as shown in Fig. 2(a),
in China, to explore the engagement of suicide attempters with to classify suicidal intention indicators according to Beck’s
intervention by direct messages. suicide intent scale.
2) General Text Features: Another direction of feature
A. Content Analysis engineering is to extract features from unstructured text. The
Users’ post on social websites reveals rich information and main features consist of N-gram features, knowledge-based
their language preferences. Through exploratory data analysis features, syntactic features, context features, and class-specific
on the user-generated content, we can have an insight into features [44]. Abboute et al. [45] built a set of keywords
language usage and linguistic clues of suicide attempters. The for vocabulary feature extraction within nine suicidal topics.
detailed analysis includes lexicon-based filtering, statistical Okhapkina et al. [46] built a dictionary of terms about suicidal
linguistic features, and topic modeling within suicide-related content. They introduced term frequency-inverse document
posts. frequency (TF-IDF) matrices for messages and a singular
Suicide-related keyword dictionary and lexicon are manu- value decomposition (SVD) for matrices. Mulholland and
ally built to enable keyword filtering [29], [30] and phrases Quinn [47] extracted vocabulary and syntactic features to build
filtering [31]. Suicide-related keywords and phrases include a classifier to predict the likelihood of a lyricist’s suicide.
“kill,” “suicide,” “feel alone,” “depressed,” and “cutting Huang et al. [48] built a psychological lexicon dictionary
myself.” Vioulès et al. [3] built a pointwise mutual information by extending HowNet (a commonsense word collection) and
symptom lexicon using an annotated Twitter data set. Gunn used a support vector machine (SVM) to detect cybersuicide
and Lester [32] analyzed posts from Twitter in the 24 h in Chinese microblogs. The topic model [49] is incorporated
before the death of a suicide attempter. Coppersmith et al. [33] with other machine learning techniques for identifying suicide
analyzed the language usage of data from the same platform. in Sina Weibo. Ji et al. [17] extracted several informative sets
Suicidal thoughts may involve strong negative feelings, anxi- of features, including statistical, syntactic, linguistic inquiry
ety, hopelessness, or other social factors, such as family and and word count (LIWC), word embedding, and topic features,
friends. Ji et al. [17] performed word cloud visualization and and then put the extracted features into classifiers, as shown
topics modeling over suicide-related content and found that in Fig. 2(b), where four traditional supervised classifiers are
suicide-related discussion covers personal and social issues. compared. Shing et al. [13] extracted several features as a bag
Colombo et al. [34] analyzed the graphical characteristics of of words (BoWs), empath, readability, syntactic features, topic
connectivity and communication in the Twitter social network. model posteriors, word embeddings, LIWC, emotion features,
Coppersmith et al. [35] provided an exploratory analysis of and mental disease lexicon.
language patterns and emotions on Twitter. Other methods Models for SID with feature engineering include SVM [44],
and techniques include Google Trends analysis for suicide risk ANNs [50], and conditional random field (CRF) [51].
monitoring [36], the reply bias assessment through linguistic Tai and Chiu [50] selected several features, including the
clues [37], human–machine hybrid method for analysis of history of suicide ideation and self-harm behavior, religious
the language effect of social support on suicidal ideation belief, family status, mental disorder history of candidates, and

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

Fig. 2. Illustrations of methods with feature engineering. (a) Neural network

with feature engineering. (b) Classifier with feature engineering.

their family. Pestian et al. [52] compared the performance of

different multivariate techniques with features of word counts,
POS, concepts, and readability scores. Similarly, Ji et al. [17]
compared four classification methods of logistic regression,
random forest, gradient boosting decision tree, and XGBoost.
Braithwaite et al. [53] validated that machine learning algo-
rithms can effectively identify high suicidal risk.
3) Affective Characteristics: Affective characteristics are
among the most distinct differences between those who
attempt suicide and healthy individuals, which has drawn con-
siderable attention from both computer scientists and mental
Fig. 3. DNNs for SID. (a) CNN. (b) RNN. (c) BERT.
health researchers. To detect the emotions in suicide notes,
Liakata et al. [51] used manual emotion categories, including
anger, sorrow, hopefulness, happiness/peacefulness, fear, pride, sophisticated feature engineering techniques. At the same
abuse, and forgiveness. Wang et al. [44] employed combined time, some also take extracted features into deep neural
characteristics of both factual (two categories) and sentimental networks (DNNs); for example, Nobles et al. [54] fed psy-
aspects (13 categories) to discover fine-grained sentiment cholinguistic features and word occurrence into the multilayer
analysis. Similarly, Pestian et al. [52] identified emotions of perceptron (MLP). Popular DNNs include convolutional neural
abuse, anger, blame, fear, guilt, hopelessness, sorrow, for- networks (CNNs), recurrent neural networks (RNNs), and bidi-
giveness, happiness, peacefulness, hopefulness, love, pride, rectional encoder representations from transformers (BERT),
thankfulness, instructions, and information. Ren et al. [14] as shown in Fig. 3(a)–(c). Natural language text is usually
proposed a complex emotion topic model and applied it to embedded into distributed vector space with popular word
analyze accumulated emotional traits in suicide blogs and embedding techniques, such as word2vec [55] and GloVe [56].
to detect suicidal intentions from a blog stream. Especially, Shing et al. [13] applied user-level CNN with the filter size
the authors studied accumulate emotional traits, including of 3, 4, and 5 to encode users’ posts. The long short-term mem-
emotion accumulation, emotion covariance, and emotion tran- ory (LSTM) network, a popular variant of RNN, is applied to
sition among eight basic emotions of joy, love, expectation, encode textual sequences and then process for classification
surprise, anxiety, sorrow, anger, and hate with a five-level with fully connected layers [17].
intensity. Recent methods introduce other advanced learning para-
digms to integrate with DNNs for SID. Ji et al. [57] proposed
model aggregation methods for updating neural networks,
C. Deep Learning i.e., CNNs and LSTMs, targeting to detect suicidal ideation in
Deep learning has been a great success in many applica- private chatting rooms. However, decentralized training relies
tions, including computer vision, NLP, and medical diagnosis. on coordinators in chatting rooms to label user posts for
In the field of suicide research, it is also an impor- supervised training, which can only be applied to minimal
tant method for automatic SID and suicide prevention. It scenarios. One possible better way is to use unsupervised or
can effectively learn text features automatically without semisupervised learning methods. Benton et al. [16] predicted

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

JI et al.: SID: A REVIEW OF MACHINE LEARNING METHODS AND APPLICATIONS 5

suicide attempt and mental health with neural models under from [69], Fig. 4(c) shows a suicide note from a website,5 and
the framework of multitask learning by predicting the gender Fig. 4(d) shows a tweet and its corresponding comments from
of users as an auxiliary task. Gaur et al. [58] incorporated Twitter.com. Nobles et al. [54] identified suicide risk using
external knowledge bases and suicide-related ontology into text messages. Some researchers also developed softwares for
a text representation and gained an improved performance suicide prevention. Berrouiguet et al. [70] developed a mobile
with a CNN model. Coppersmith et al. [59] developed a application for health status self-report. Meyer et al. [71]
deep learning model with GloVe for word embedding, bidi- developed an e-PASS Suicidal Ideation Detector (eSID) tool
rectional LSTM for sequence encoding, and self-attention for medical practitioners. Shah et al. [72] utilized social media
mechanism for capturing the most informative subsequence. videos and studied multimodal behavioral markers.
Sawhney et al. [60] used LSTM, CNN, and RNN for SID.
Similarly, Tadesse et al. [61] employed the LSTM-CNN
model. Ji et al. [62] proposed an attentive relation network A. Questionnaires
with LSTM and topic modeling for encoding text and risk Mental disorder scale criteria, such as DSM-IV,6
indicators. ICD-10,7 and the IPDE-SQ, provides good tool for evaluating
In the 2019 CLPsych Shared Task [63], many popular an individual’s mental status and their potential for suicide.
DNN architectures were applied. Hevia et al. [64] evaluated Those criteria and examination metrics can be used to
the effect of pretraining using different models, including design questionnaires for self-measurement or face-to-face
GRU-based RNN. Morales et al. [65] studied several popular clinician-patient interview.
deep learning models, such as CNN, LSTM, and Neural To study the assessment of suicidal behavior,
Network Synthesis (NeuNetS). Matero et al. [66] proposed Delgado-Gomez et al. [10] applied and compared the
a dual-context model using hierarchically attentive RNN and IPDE-SQ and the “Barrat’s Impulsiveness Scale” (version 11,
BERT. BIS-11) to identify people likely to attempt suicide. The
Another subdirection is the so-called hybrid method that authors also conducted a study on individual items from those
cooperates minor feature engineering with representation two scales. The BIS-11 scale has 30 items with four-point
learning techniques. Chen et al. [67] proposed a hybrid ratings, while the IPDE-SQ in DSM-IV has 77 true-false
classification model of the behavioral model and the suicide screening questions. Furthermore, Delgado-Gomez et al. [42]
language model. Zhao et al. [68] proposed the D-CNN model introduced the “Holmes–Rahe SRRS” and the IPDE-SQ as
taking word embedding and external tabular features as inputs well to two comparison groups of suicide attempters and
for classifying suicide attempters with depression. nonsuicide attempters. The SRRS consists of 43 ranked
life events of different levels of severity. Harris et al. [73]
D. Summary surveyed understanding suicidal individuals’ online behaviors
to assist suicide prevention. Sueki [74] conducted an online
The popularization of machine learning has facilitated
panel survey among Internet users to study the association
research on SID from multimodal data and provided a promis-
between suicide-related Twitter use and suicidal behavior.
ing way for effective early warning. Current research focuses
Based on the questionnaire results, they applied several
on text-based methods by extracting features and deep learning
supervised learning methods, including linear regression,
for automatic feature learning. Researchers widely use many
stepwise linear regression, decision trees, Lars-en, and SVMs,
canonical NLP features, such as TF-IDF, topics, syntactic,
to classify suicidal behaviors.
affective characteristics, readability, and deep learning mod-
els, such as CNN and LSTM. Those methods, especially
DNNs with automatic feature learning, boosted predictive B. Electronic Health Records
performance and preliminary success on suicidal intention
understanding. However, some methods may only learn sta- The increasing volume of EHRs has paved the way for
tistical cues and lack of commonsense. The recent work [58] machine learning techniques for suicide attempter predic-
incorporated external knowledge using knowledge bases and tion. Patient records include demographical information and
suicide ontology for knowledge-aware suicide risk assessment. diagnosis-related history, such as admissions and emergency
It took a remarkable step toward knowledge-aware detection. visits. However, due to the data characteristics, such as spar-
sity, variable length of clinical series, and heterogeneity of
patient records, many challenges remain in modeling medical
III. A PPLICATIONS ON D OMAINS
data for suicide attempt prediction. Besides, the recording
Many machine learning techniques have been introduced procedures may change because of the change of healthcare
for SID. The relevant extant research can also be viewed policies and the update of diagnosis codes.
according to the data source. Specific applications cover a There are several works of predicting suicide risk based
wide range of domains, including questionnaires, electronic on EHRs [75], [76]. Tran et al. [69] proposed an integrated
health records (EHRs), suicide notes, and online user content. suicide risk prediction framework with a feature extraction
Fig. 4 shows some examples of data source for SID, where
Fig. 4(a) lists selected questions of the “International Personal 5 https://paranorms.com/suicide-notes
Disorder Examination Screening Questionnaire” (IPDE-SQ) 6 https://psychiatry.org/psychiatrists/practice/dsm
adapted from [42], Fig. 4(b) shows selected patient’s records 7 https://apps.who.int/classifications/icd10/browse/2016/en

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

Fig. 4. Examples of content for SID. (a) Questionnaire. (b) EHR. (c) Suicide notes. (d) Tweets.

TABLE I
C ATEGORIZATION OF M ETHODS FOR SID

scheme, risk classifiers, and risk calibration procedure. Explic- medium-term suicide risk assessments. By introducing random
itly, each patient’s clinical history is represented as a tem- forests, gradient boosting machines, and DNNs, the authors
poral image. Iliou et al. [77] proposed a data preprocessing managed to deal with high dimensionality and redundancy
method to boost machine learning techniques for suicide ten- issues of data. Although the previous method gained prelimi-
dency prediction of patients suffering from mental disorders. nary success, Iliou et al. [77] and Nguyen et al. [78] have a
Nguyen et al. [78] explored real-world administrative data limitation on the source of data, which focuses on patients
of mental health patients from the hospital for short- and with mental disorders in their historical records. Bhat and

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

JI et al.: SID: A REVIEW OF MACHINE LEARNING METHODS AND APPLICATIONS 7

Goldman-Mellor [79] used an anonymized general EHR data data analysis for contents that suggests suicidal tendencies in
set to relax the restriction on patient’s diagnosis-related history Twitter social networks. The authors used the data from an
and applied neural networks as a classification model to predict online conversation called #dearmentalhealthprofessionals.
suicide attempters. Another famous platform Reddit is an online forum with
topic-specific discussions has also attracted much research
C. Suicide Notes interest for studying mental health issues [89] and suicidal
ideation [37]. A community on Reddit called SuicideWatch
Suicide notes are the written notes left by people before
is intensively used for studying suicidal intention [17], [90].
committing suicide. They are usually written on letters and
De Choudhury et al. [90] applied a statistical methodology to
online blogs and recorded in audio or video. Suicide notes
discover the transition from mental health issues to suicidality.
provide material for NLP research. Previous approaches have
Kumar et al. [91] examined the posting activity following the
examined suicide notes using content analysis [52], senti-
celebrity suicides, studied the effect of celebrity suicides on
ment analysis [44], [80], and emotion detection [51]. Pes-
suicide-related contents, and proposed a method to prevent the
tian et al. [52] used transcribed suicide notes with two
high-profile suicides.
groups of completers and elicitors from people who have a
Many pieces of research [48], [49] work on detecting
personality disorder or potential morbid thoughts. White and
suicidal ideation in Chinese microblogs. Guan et al. [92]
Mazlack [81] analyzed word frequencies in suicide notes using
studied user profile and linguistic features for estimating
a fuzzy cognitive map to discern causality. Liakata et al. [51]
suicide probability in Chinese microblogs. There also remains
employed machine learning classifiers to 600 suicide messages
some work using other platforms for SID. For example,
with varied length, different readability quality, and multiclass
Cash et al. [93] conducted a study on adolescents’ comments
annotations.
and content analysis on MySpace. Steaming data provides
Emotion in text provides sentimental cues of suici-
a good source for user pattern analysis. Vioulès et al. [3]
dal ideation understanding. Desmet and Hoste [82] con-
conducted user- and post-centric behavior analysis and applied
ducted a fine-grained emotion detection on suicide notes
a martingale framework to detect sudden emotional changes in
of 2011 i2b2 task. Wicentowski and Sydes [83] used an
the Twitter data stream for monitoring suicide warning signs.
ensemble of maximum entropy classification. Wang et al. [44]
Ren et al. [14] used the blog stream collected from public blog
and Kovačević et al. [84] proposed hybrid machine learning
articles written by suicide victims to study the accumulated
and rule-based method for the i2b2 sentiment classification
emotional information.
task in suicide notes.
In the age of cyberspace, more suicide notes are now written
in the form of web blogs and can be identified as carrying the E. Summary
potential risk of suicide. Huang et al. [29] monitored online Applications of SID mainly consist of four domains: ques-
blogs from MySpace.com to identify at-risk bloggers. Schoene tionnaires, EHRs, suicide notes, and online user content.
and Dethlefs [85] extracted linguistic and sentiment features Table II gives a summary of categories, data sources, and
to identity genuine suicide notes and comparison corpus. methods. Among these four main domains, questionnaires
and EHRs require self-report measurement or patient–clinician
D. Online User Content interactions and rely highly on social workers or mental health
The widespread use of mobile Internet and social network- professions. Suicide notes have a limitation on immediate
ing services facilitates people’s expressing their life events prevention, as many suicide attempters commit suicide in a
and feelings freely. As social websites provide an anonymous short time after they write suicide notes. However, they provide
space for online discussion, an increasing number of people a good source for content analysis and the study of suicide
suffering from mental disorders turn to seek for help. There factors. The last online user content domain is one of the most
is a concerning tendency that potential suicide victims post promising ways of early warning and suicide prevention when
their suicidal thoughts on social websites, such as Facebook, empowered with machine learning techniques. With the rapid
Twitter, Reddit, and MySpace. Social media platforms are development of digital technology, user-generated content will
becoming a promising tunnel for monitoring suicidal thoughts play a more important role in SID. Other forms of data, such as
and preventing suicide attempts [86]. Massive user-generated health data generated by wearable devices, can be very likely
data provide a good source to study online users’ language to help with suicide risk monitoring in the future.
patterns. Using data mining techniques on social networks
and applying machine learning techniques provide an avenue IV. TASKS AND DATA S ETS
to understand the intent within online posts, provide early In this section, we summarize specific tasks in SID and
warnings, and even relieve a person’s suicidal intentions. other suicide-related tasks about mental disorders. Some tasks,
Twitter provides a good source for research on suicidality. such as reasoning suicidal messages, generating a response,
O’Dea et al. [12] collected tweets using the public API and and suicide attempters’ detection on a social graph, may
developed automatic suicide detection by applying logistic lack benchmarks for evaluation. However, they are critical for
regression and SVM on TF-IDF features. Wang et al. [87] effective detection. We propose these tasks together with the
further improved the performance with effective feature engi- current research direction and call for contribution to these
neering. Shepherd et al. [88] conducted psychology-based tasks from the research community. Meanwhile, an elaborate

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

For Experiment
TABLE II mental pain, and cognitive attribution aspects associated with
S UMMARY OF S TUDIES ON SID F ROM THE V IEWS OF I NTERVENTION suicide.
C ATEGORIES , D ATA , AND M ETHODS
3) Suicide Attempter Detection: The two tasks mentioned
earlier focus on a single text itself. However, the primary pur-
pose of SID is the identify suicide attempters. Thus, it is vital
to achieving user-level detection, which consists of two folds,
i.e., user-level multiinstance suicidality detection and suicide
attempt detection on a graph. The former takes a bag of posts
from individuals as input and conducts multiinstance learning
over a bag of messages. The later identifies suicide attempters
in a specific social graph built by the interaction between users
in social networks. It considers the relationship between social
users and can be regarded as a node classification problem in
a graph.
4) Generating Response: The ultimate goal of SID is inter-
vention and suicide prevention. Many people with suicidal
intentions tend to post their suffering at midnight. Another task
is generating a thoughtful response for counseling potential
suicidal victims to enable immediate social care and relieve
their suicidal intention. Gilat et al. [94] introduced eight types
list of data sets for currently available tasks is provided, and
of response strategies; they are emotional support, offering
some potential data sources are also described to promote the
group support, empowerment, interpretation, cognitive change
research efforts.
inducement, persuasion, advising, and referring. This task
requires machine learning techniques, especially sequence-
A. Tasks to-sequence learning, to have the ability to adopt effective
1) Suicide Text Classification: The first task—suicide text response strategies to generate better response and eliminate
classification—can be viewed as a domain-specific application people’s suicidality. When social workers or volunteers go
of general text classification, which includes binary and mul- back online, this response generation technique can also gen-
ticlass classification. Binary suicidality classification simply erate hints for them to compose a thoughtful response.
determines text with suicidal ideation or not, while multiclass 5) Mental Disorders and Self-Harm Risk: Suicidal ideation
suicidality classification conducts fine-grained suicide risk has a strong relationship with a mental health issue and
assessment. For example, some studies divide suicide risk self-harm risks. Thus, detecting severe mental disorders or
into four levels: no, low, moderate, and severe. Alternatively, self-harm risks is also an important task. Such works include
it can also consider four types of class labels according to depression detection [95], self-harm detection [96], stressful
mental and behavioral procedures, i.e., nonsuicidal, suicidal periods and stressor events detection [97], building knowledge
thoughts/wishes, suicidal intentions, and suicidal act/plan. graph for depression [98], and correlation analysis on depres-
Another subtask is risk assessment by learning from multi- sion and anxiety [99]. Corresponding subtasks in this field are
aspect suicidal posts. Adopting the definition of characteristics similar to suicide text classification in Section IV-A1.
of suicidal messages, Gilat et al. [94] manually tagged suicidal
posts with multiaspect labels, including mental pain, cognitive B. Data Sets
attribution, and level of suicidal risk. Mental pain includes 1) Text Data:
the loss of control, acute loneliness, emptiness, narcissistic a) Reddit: Reddit is a registered online community that
wounds, irreversibility loss of energy, and emotional flooding, aggregates social news and online discussions. It consists of
scaled into [0, 7]. Cognitive attribution is the frustration of many topic categories, and each area of interest within a topic
needs associated with interpersonal relationships, or there is is called a subreddit. A subreddit called “Suicide Watch”
no indication of attribution. (SW)8 is intensively used for further annotation as positive
2) Reasoning Suicidal Messages: Massive data mining and samples. Posts without suicidal content are sourced from other
machine learning algorithms have achieved remarkable out- popular subreddits. Ji et al. [17] released a data set with
comes by using DNNs. However, simple feature sets and 3549 posts with suicidal ideation. Shing et al. [13] published
classification models are not predictive enough to detect their UMD Reddit Suicidality data set with 11 129 users
complicated suicidal intentions. Machine learning techniques and total 1 556 194 posts and sampled 934 users for further
require reasoning suicidal messages to have a more in-depth annotation. Aladağ et al. [100] collected 508 398 posts using
insight into suicidal factors and the innermost being from Google Cloud BigQuery and manually annotated 785 posts.
textual posts. This task aims to employ interpretable meth- b) Twitter: Twitter is a popular social networking service,
ods to investigate suicidal factors and incorporate them with where many users also talk about their suicidal ideation. Twit-
commonsense reasoning, which may improve the prediction of ter is quite different from Reddit in post length, anonymity,
suicidal factors. Specific tasks include automatic summariza-
tion of suicide factor, finding an explanation of suicidal risk in 8 https://reddit.com/r/SuicideWatch

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

JI et al.: SID: A REVIEW OF MACHINE LEARNING METHODS AND APPLICATIONS 9

and the way communication and interaction. Twitter user DNN-based representation learning techniques. However, cur-
data with suicidal ideation and depression are collected by rent research has several limitations, and there are still great
Coppersmith et al. [33]. Ji et al. [17] collected an imbalanced challenges for future work.
data set of 594 tweets with suicidal ideation out of a total
of 10 288 tweets. Vioulès et al. collected 5446 tweets using A. Limitations
Twitter streaming API [3], of which 2381 and 3065 tweets
1) Data Deficiency: The most critical issue of current
are from the distressed users and normal users, respectively.
research is data deficiency. Current methods mainly apply
However, most Twitter-based data sets are no longer available
supervised learning techniques that require manual annotation.
as per the policy of Twitter.
However, there are not enough annotated data to support
c) ReachOut: ReachOut Forum9 is a peer support plat-
further research. For example, labeled data with fine-grained
form provided by an Australian mental health care organiza-
suicide risk only have limited instances, and there are no
tion. The ReachOut data set [101] was first released in the
multiaspect data and data with social relationships.
CLPsych17 shared task. Participants were initially given a
2) Annotation Bias: There is little evidence to confirm the
training data set of 65 756 forum posts, of which 1188 were
suicide action to obtain ground truth. Thus, current data are
annotated manually with the expected category, and a test
obtained by manual labeling with some predefined annotation
set of 92 207 forum posts, of which 400 were identified as
rules. The crowdsourcing-based annotation may lead to bias
requiring annotation. The specific four categories are described
of labels. Shing et al. [13] asked experts for labeling but
as follows.
only obtained a limited number of labeled instances. As
1) Crisis: The author or someone else is at risk of harm. for the demographical data, the quality of suicide data is
2) Red: The post should be responded to as soon as concerning, and mortality estimation is general death but not
possible. suicide.11 Some cases are misclassified as accidents or death
3) Amber: The post should be responded to at some point of undetermined intent.
if the community does not rally strongly around it. 3) Data Imbalance: Posts with suicidal intention account
4) Green: The post can be safely ignored or left for the for a tiny proportion of massive social posts. However, most
community to address. works built data sets in an approximately even manner to
2) EHR: EHR data contain demographical information, collect relatively balanced positive and negative samples rather
admissions, diagnostic reports, and physician notes. A col- than treating it as an ill-balanced data distributed.
lection of EHRs is from the California emergency department 4) Lack of Intention Understanding: The current statistical
encounter and hospital admission. It contains 522 056 anony- learning method failed to have a good understanding of
mous EHR records from California-resident adolescents. How- suicidal intention. The psychology behind suicidal attempts
ever, it is not public for access. Bhat and Goldman-Mellor [79] is complex. However, mainstream methods focus on selecting
first used these records from 2006 to 2009 to predict the features or using complex neural architectures to boost the
suicide attempt in 2010. Haerian et al. [102] selected 280 cases predictive performance. From the phenomenology of suicidal
for evaluation from the clinical data warehouse (CDW) posts in social content, machine learning methods learned
and WebCIS database at the NewYork Presbyterian Hos- statistical clues. However, they failed to reason over the risk
pital/Columbia University Medical Center. Tran et al. [69] factors by incorporating the psychology of suicide.
studied emergency attendances with a least one risk assessment
from the Barwon Health data warehouse. The selected data set B. Future Work
contains 7746 patients and 17 771 assessments.
1) Emerging Learning Techniques: The advances of deep
3) Mental Disorders: Mental health issues, such as depres-
learning techniques have boosted research on SID. More
sion without effective treatment, can turn into suicidal ideation.
emerging learning techniques, such as attention mechanism
For the convenience of research on mental disorders, we also
list several resources for monitoring mental disorders. The and graph neural networks, can be introduced for suicide
text representation learning. Other learning paradigms, such
eRisk data set of early detection of signs of depression [103]
as transfer learning, adversarial training, and reinforcement
is released by the first task of the 2018 workshop at the
learning, can also be utilized. For example, knowledge of the
Conference and Labs of the Evaluation Forum (CLEF), which
mental health detection domain can be transferred for SID,
focuses on early risk prediction on the Internet.10 This data set
and generative adversarial networks can be used to generate
contains sequential text from social media. Another data set is
adversarial samples for data augmentation.
the Reddit Self-reported Depression Diagnosis (RSDD) data
In social networking services, posts with suicidal ideation
set [95], which contains 9000 diagnosed users with depression
and approximately 107 000 matched control users. are in the long tail of the distribution of different post
categories. To achieve effective detection in the ill-balanced
V. D ISCUSSION AND F UTURE W ORK distribution of real-world scenarios, few-shot learning can be
utilized to train on a few labeled posts with suicidal ideation
Many preliminary works have been conducted for SID, among the large social corpus.
especially boosted by manual feature engineering and
11 World Health Organization, Preventing suicide: a global imperative, 2014:
9 https://au.reachout.com/forums
https://apps.who.int/iris/bitstream/handle/10665/131056/9789241564779_eng.
10 https://early.irlab.org pdf.

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

TABLE III
S UMMARY OF THE P UBLIC D ATA S ETS

2) Suicidal Intention Understanding and Interpretability: comments, or get their suicidality relieved. A score will be
Many factors are correlated with suicide, such as mental defined by observing the reaction from a suicide attempter as
health, economic recessions, gun prevalence, daylight patterns, a reward. The conversational suicide intervention uses a policy
divorce laws, media coverage of suicide, and alcohol use.12 gradient for agents to generated responses with maximum
A better understanding of suicidal intention can provide a rewards to best relieve people’s suicidal thoughts.
guideline for effective detection and intervention. A new
research direction is to equip deep learning models with com- VI. C ONCLUSION
monsense reasoning, for example, by incorporating external
suicide-related knowledge bases. Suicide prevention remains an essential task in our modern
Deep learning techniques can learn an accurate prediction society. Early detection of suicidal ideation is an important
model. However, this would be a black-box model. In order and effective way to prevent suicide. This survey investi-
to better understand people’s suicidal intentions and have gates existing methods for SID from a broad perspective that
a reliable prediction, new interpretable models should be covers clinical methods, such as patient–clinician interaction
developed. and medical signal sensing; textual content analysis, such as
3) Temporal Suicidal Ideation Detection: Another direction lexicon-based filtering and word cloud visualization; feature
is to detect suicidal ideation over the data stream and consider engineering, including tabular, textual, and affective features;
the temporal information. There exist several stages of suicide and deep learning-based representation learning, such as CNN-
attempts, including stress, depression, suicidal thoughts, and and LSTM-based text encoders. Four main domain-specific
suicidal plan. Modeling people’s posts’ temporal trajectory can applications on questionnaires, EHRs, suicide notes, and
effectively monitor the change of mental status and is essential online user content are introduced.
for detecting early signs of suicidal ideation. Psychological experts have conducted most work in
this field with statistical analysis and computer scientists
4) Proactive Conversational Intervention: The ultimate aim
with feature engineering-based machine learning and deep
of SID is intervention and prevention. Very little work is
learning-based representation learning. Based on current
undertaken to enable proactive intervention. Proactive suicide
research, we summarized existing tasks and further proposed
prevention online (PSPO) [105] provides a new perspective
new possible tasks. Last but not least, we discuss some
with the combination of suicidal identification and crisis
limitations of current research and propose a series of future
management. An effective way is through conversations. Auto-
directions, including utilizing emerging learning techniques,
matic response generation becomes a promising technical
interpretable intention understanding, temporal detection, and
solution to enable timely intervention for suicidal thoughts.
proactive conversational intervention.
Natural language generation techniques can be utilized to
Online social content is very likely to be the main channel
generate counseling responses to comfort people’s depres-
for SID in the future. Therefore, it is essential to develop new
sion or suicidal ideation. Reinforcement learning can also be
methods, which can heal the schism between clinical mental
applied for conversational suicide intervention. After suicide
health detection and automatic machine detection, to detect
attempters post suicide messages (as the initial state), online
online texts containing suicidal ideation in the hope that
volunteers and lay individuals will take action to comment
suicide can be prevented.
on the original posts and persuade attempters to give up
their suicidality. The attempter may do nothing, reply to the
R EFERENCES
12 Report by Lindsay Lee, Max Roser, and Esteban Ortiz-Ospina in Our- [1] S. Hinduja and J. W. Patchin, “Bullying, cyberbullying, and suicide,”
WorldInData.org, retrieved from https://ourworldindata.org/suicide. Arch. Suicide Res., vol. 14, no. 3, pp. 206–221, Jul. 2010.

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

JI et al.: SID: A REVIEW OF MACHINE LEARNING METHODS AND APPLICATIONS 11

[2] J. Joo, S. Hwang, and J. J. Gallo, “Death ideation and suicidal ideation [24] D. Sikander et al., “Predicting risk of suicide using resting state heart
in a community sample who do not meet criteria for major depression,” rate,” in Proc. Asia–Pacific Signal Inf. Process. Assoc. Annu. Summit
Crisis, vol. 37, no. 2, pp. 161–165, Mar. 2016. Conf. (APSIPA), Dec. 2016, pp. 1–4.
[3] M. J. Vioules, B. Moulahi, J. Aze, and S. Bringay, “Detection of [25] M. A. Just et al., “Machine learning of neural representations of suicide
suicide-related posts in Twitter data streams,” IBM J. Res. Develop., and emotion concepts identifies suicidal youth,” Nature Hum. Behav.,
vol. 62, no. 1, pp. 7:1–7:12, Jan. 2018. vol. 1, no. 12, pp. 911–919, 2017.
[4] A. J. Ferrari et al., “The burden attributable to mental and substance [26] N. Jiang, Y. Wang, L. Sun, Y. Song, and H. Sun, “An ERP study of
use disorders as risk factors for suicide: Findings from the global implicit emotion processing in depressed suicide attempters,” in Proc.
burden of disease study 2010,” PLoS ONE, vol. 9, no. 4, Apr. 2014, 7th Int. Conf. Inf. Technol. Med. Educ. (ITME), Nov. 2015, pp. 37–40.
Art. no. e91936. [27] M. Lotito and E. Cook, “A review of suicide risk assessment instru-
[5] R. C. O’Connor and M. K. Nock, “The psychology of suicidal ments and approaches,” Mental Health Clinician, vol. 5, no. 5,
behaviour,” Lancet Psychiatry, vol. 1, no. 1, pp. 73–85, 2014. pp. 216–223, Sep. 2015.
[6] J. Lopez-Castroman et al., “Mining social networks to improve suicide [28] Z. Tan, X. Liu, X. Liu, Q. Cheng, and T. Zhu, “Designing microblog
prevention: A scoping review,” J. Neurosci. Res., vol. 98, no. 4, direct messages to engage social media users with suicide ideation:
pp. 616–625, Apr. 2020. Interview and survey study on weibo,” J. Med. Internet Res., vol. 19,
[7] C. M. McHugh, A. Corderoy, C. J. Ryan, I. B. Hickie, and no. 12, p. e381, Dec. 2017.
M. M. Large, “Association between suicidal ideation and suicide: Meta- [29] Y.-P. Huang, T. Goh, and C. L. Liew, “Hunting suicide notes in Web
analyses of odds ratios, sensitivity, specificity and positive predictive 2.0–preliminary findings,” in Proc. 9th IEEE Int. Symp. Multimedia
value,” BJPsych Open, vol. 5, no. 2, Mar. 2019. Workshops (ISMW), Dec. 2007, pp. 517–521.
[8] G. Kassen, A. Kudaibergenova, A. Mukasheva, D. Yertargynkyzy, and [30] K. D. Varathan and N. Talib, “Suicide detection system based on
K. Moldassan, “Behavioral risk factors for suicide among adoles- Twitter,” in Proc. Sci. Inf. Conf., Aug. 2014, pp. 785–788.
cent schoolchildren,” Elementary Educ. Online, vol. 19, pp. 66–77, [31] J. Jashinsky et al., “Tracking suicide risk factors through Twitter in the
Jan. 2020. US,” Crisis, vol. 35, no. 1, pp. 51–59, Jan. 2014.
[9] V. Venek, S. Scherer, L.-P. Morency, A. S. Rizzo, and J. Pestian, [32] J. F. Gunn and D. Lester, “Twitter postings and suicide: An analysis
“Adolescent suicidal risk assessment in clinician-patient interaction,” of the postings of a fatal suicide in the 24 hours prior to death,”
IEEE Trans. Affect. Comput., vol. 8, no. 2, pp. 204–215, Apr. 2017. Suicidologi, vol. 17, no. 3, pp. 28–30, Jun. 2015.
[10] D. Delgado-Gomez, H. Blasco-Fontecilla, A. A. Alegria, [33] G. Coppersmith, R. Leary, E. Whyne, and T. Wood, “Quantifying
T. Legido-Gil, A. Artes-Rodriguez, and E. Baca-Garcia, “Improving suicidal ideation via language usage on social media,” in Proc. Joint
the accuracy of suicide attempter classification,” Artif. Intell. Med., Statist. Meetings Stat. Comput. Sect. (JSM), 2015, pp. 1–15.
vol. 52, no. 3, pp. 165–168, Jul. 2011. [34] G. B. Colombo, P. Burnap, A. Hodorog, and J. Scourfield, “Analysing
[11] G. Liu, C. Wang, K. Peng, H. Huang, Y. Li, and W. Cheng, “SocInf: the connectivity and communication of suicidal users on Twitter,”
Membership inference attacks on social media health data with machine Comput. Commun., vol. 73, pp. 291–300, Jan. 2016.
learning,” IEEE Trans. Comput. Social Syst., vol. 6, no. 5, pp. 907–921, [35] G. Coppersmith, K. Ngo, R. Leary, and A. Wood, “Exploratory analysis
Oct. 2019. of social media prior to a suicide attempt,” in Proc. 3rd Workshop
Comput. Lingusitics Clin. Psychol., 2016, pp. 106–117.
[12] B. O’Dea, S. Wan, P. J. Batterham, A. L. Calear, C. Paris, and
[36] P. Solano et al., “A Google-based approach for monitoring suicide risk,”
H. Christensen, “Detecting suicidality on Twitter,” Internet Intervent.,
Psychiatry Res., vol. 246, pp. 581–586, Dec. 2016.
vol. 2, no. 2, pp. 183–188, May 2015.
[37] H. Y. Huang and M. Bashir, “Online community and suicide prevention:
[13] H.-C. Shing, S. Nair, A. Zirikly, M. Friedenberg, H. Daumé, III, and
Investigating the linguistic cues and reply bias,” in Proc. CHI Conf.
P. Resnik, “Expert, crowdsourced, and machine assessment of suicide
Hum. Factors Comput. Syst., 2016, pp. 1–5.
risk via online postings,” in Proc. 5th Workshop Comput. Linguistics
[38] M. De Choudhury and E. Kıcıman, “The language of social support
Clin. Psychol., Keyboard Clinic, 2018, pp. 25–36.
in social media and its effect on suicidal ideation risk,” in Proc. 11th
[14] F. Ren, X. Kang, and C. Quan, “Examining accumulated emotional Int. AAAI Conf. Web Social Media, 2017, p. 32.
traits in suicide blogs with an emotion topic model,” IEEE J. Biomed.
[39] M. E. Larsen et al., “The use of technology in suicide prevention,”
Health Informat., vol. 20, no. 5, pp. 1384–1396, Sep. 2016.
in Proc. 37th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC),
[15] L. Yue, W. Chen, X. Li, W. Zuo, and M. Yin, “A survey of senti- Aug. 2015, pp. 7316–7319.
ment analysis in social media,” Knowl. Inf. Syst., vol. 60, pp. 1–47, [40] N. Masuda, I. Kurahashi, and H. Onari, “Suicide ideation of individuals
Aug. 2018. in online social networks,” PLoS ONE, vol. 8, no. 4, Apr. 2013,
[16] A. Benton, M. Mitchell, and D. Hovy, “Multi-task learning for mental Art. no. e62262.
health using social media text,” in in Proc. EACL. Stroudsburg, PA, [41] S. Chattopadhyay, “A study on suicidal risk analysis,” in Proc. 9th Int.
USA: Association for Computational Linguistics, 2017, pp. 152–162. Conf. e-Health Netw., Appl. Services, Jun. 2007, pp. 74–78.
[17] S. Ji, C. P. Yu, S.-F. Fung, S. Pan, and G. Long, “Supervised learning [42] D. Delgado-Gomez, H. Blasco-Fontecilla, F. Sukno, M. Socorro
for suicidal ideation detection in online user content,” Complexity, Ramos-Plasencia, and E. Baca-Garcia, “Suicide attempters classifica-
vol. 2018, pp. 1–10, Sep. 2018. tion: Toward predictive models of suicidal behavior,” Neurocomputing,
[18] S. Ji, G. Long, S. Pan, T. Zhu, J. Jiang, and S. Wang, “Detecting vol. 92, pp. 3–8, Sep. 2012.
suicidal ideation with data protection in online communities,” in [43] S. Chattopadhyay, “A mathematical model of suicidal-intent-estimation
Proc. 24th Int. Conf. Database Syst. Adv. Appl. (DASFAA). Cham, in adults,” Amer. J. Biomed. Eng., vol. 2, no. 6, pp. 251–262, Jan. 2013.
Switzerland: Springer, 2019, pp. 225–229. [44] W. Wang, L. Chen, M. Tan, S. Wang, and A. P. Sheth, “Discovering
[19] J. Tighe, F. Shand, R. Ridani, A. Mackinnon, N. De La Mata, and fine-grained sentiment in suicide notes,” Biomed. Informat. Insights,
H. Christensen, “Ibobbly mobile health intervention for suicide pre- vol. 5, no. 1, p. 137, 2012.
vention in australian indigenous youth: A pilot randomised controlled [45] A. Abboute, Y. Boudjeriou, G. Entringer, J. Azé, S. Bringay, and
trial,” BMJ Open, vol. 7, no. 1, Jan. 2017, Art. no. e013518. P. Poncelet, “Mining Twitter for suicide prevention,” in Proc. Int. Conf.
[20] N. N. Gomes de Andrade, D. Pawson, D. Muriello, L. Donahue, and Appl. Natural Lang. Data Bases/Inf. Syst. Cham, Switzerland: Springer,
J. Guadagno, “Ethics and artificial intelligence: Suicide prevention 2014, pp. 250–253.
on Facebook,” Philosophy Technol., vol. 31, no. 4, pp. 669–684, [46] E. Okhapkina, V. Okhapkin, and O. Kazarin, “Adaptation of infor-
Dec. 2018. mation retrieval methods for identifying of destructive informational
[21] L. C. McKernan, E. W. Clayton, and C. G. Walsh, “Protecting life influence in social networks,” in Proc. 31st Int. Conf. Adv. Inf. Netw.
while preserving liberty: Ethical recommendations for suicide preven- Appl. Workshops (WAINA), Mar. 2017, pp. 87–92.
tion with artificial intelligence,” Frontiers Psychiatry, vol. 9, p. 650, [47] M. Mulholland and J. Quinn, “Suicidal tendencies: The automatic
Dec. 2018. classification of suicidal and non-suicidal lyricists using NLP,” in Proc.
[22] K. P. Linthicum, K. M. Schafer, and J. D. Ribeiro, “Machine learning IJCNLP, 2013, pp. 680–684.
in suicide science: Applications and ethics,” Behav. Sci. Law, vol. 37, [48] X. Huang, L. Zhang, D. Chiu, T. Liu, X. Li, and T. Zhu, “Detecting
no. 3, pp. 214–222, May 2019. suicidal ideation in chinese microblogs with psychological lexicons,”
[23] S. Scherer, J. Pestian, and L.-P. Morency, “Investigating the speech in Proc. IEEE 11th Int. Conf. Ubiquitous Intell. Comput. Autonomic
characteristics of suicidal adolescents,” in Proc. IEEE Int. Conf. Trusted Comput. IEEE 14th Int. Conf. Scalable Comput. Commun.
Acoust., Speech Signal Process., May 2013, pp. 709–713. Associated Workshops, Dec. 2014, pp. 844–849.

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

[49] X. Huang, X. Li, T. Liu, D. Chiu, T. Zhu, and L. Zhang, “Topic model [72] A. P. Shah, V. Vaibhav, V. Sharma, M. Al Ismail, J. Girard, and
for identifying suicidal ideation in chinese microblog,” in Proc. 29th L.-P. Morency, “Multimodal behavioral markers exploring suicidal
Pacific Asia Conf. Lang., Inf. Comput., 2015, pp. 553–562. intent in social media videos,” in Proc. Int. Conf. Multimodal Interact.,
[50] Y.-M. Tai and H.-W. Chiu, “Artificial neural network analysis on Oct. 2019, pp. 409–413.
suicide and self-harm history of taiwanese soldiers,” in Proc. 2nd Int. [73] K. M. Harris, J. P. McLean, and J. Sheffield, “Suicidal and online: How
Conf. Innov. Comput., Inf. Control (ICICIC), Sep. 2007, p. 363. do online behaviors inform us of this high-risk population?” Death
[51] M. Liakata, J. H. Kim, S. Saha, J. Hastings, and Stud., vol. 38, no. 6, pp. 387–394, Jul. 2014.
D. Rebholz-Schuhmann, “Three hybrid classifiers for the detection [74] H. Sueki, “The association of suicide-related Twitter use with suicidal
of emotions in suicide notes,” Biomed. Informat. Insights, vol. 2012, behaviour: A cross-sectional study of young Internet users in Japan,”
no. 1, pp. 175–184, 2012. J. Affect. Disorders, vol. 170, pp. 155–160, Jan. 2015.
[52] J. Pestian, H. Nasrallah, P. Matykiewicz, A. Bennett, and A. Leenaars,
[75] K. W. Hammond, R. J. Laundry, T. M. Oleary, and W. P. Jones, “Use of
“Suicide note classification using natural language processing: A text search to effectively identify lifetime prevalence of suicide attempts
content analysis,” Biomed. Inform. Insights, vol. 2010, no. 3, p. 19,
among veterans,” in Proc. 46th Hawaii Int. Conf. Syst. Sci., Jan. 2013,
2010.
pp. 2676–2683.
[53] S. R. Braithwaite, C. Giraud-Carrier, J. West, M. D. Barnes, and
C. L. Hanson, “Validating machine learning algorithms for Twitter [76] C. G. Walsh, J. D. Ribeiro, and J. C. Franklin, “Predicting risk of
data against established measures of suicidality,” JMIR Mental Health, suicide attempts over time through machine learning,” Clin. Psychol.
vol. 3, no. 2, p. e21, May 2016. Sci., vol. 5, no. 3, pp. 457–469, May 2017.
[54] A. L. Nobles, J. J. Glenn, K. Kowsari, B. A. Teachman, and [77] T. Iliou et al., “Machine learning preprocessing method for suicide
L. E. Barnes, “Identification of imminent suicide risk among young prediction,” in Artificial Intelligence Applications and Innovations,
adults using text messages,” in Proc. CHI Conf. Hum. Factors Comput. L. Iliadis and I. Maglogiannis, Eds. Cham, Switzerland: Springer, 2016,
Syst. (CHI), 2018, pp. 1–11. pp. 53–60.
[55] T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estimation [78] T. Nguyen, T. Tran, S. Gopakumar, D. Phung, and S. Venkatesh,
of word representations in vector space,” 2013, arXiv:1301.3781. “An evaluation of randomized machine learning methods for redundant
[Online]. Available: http://arxiv.org/abs/1301.3781 data: Predicting short and medium-term suicide risk from administra-
[56] J. Pennington, R. Socher, and C. Manning, “Glove: Global vectors for tive records and risk assessments,” 2016, arXiv:1605.01116. [Online].
word representation,” in Proc. Conf. Empirical Methods Natural Lang. Available: http://arxiv.org/abs/1605.01116
Process. (EMNLP), 2014, pp. 1532–1543. [79] H. S. Bhat and S. J. Goldman-Mellor, “Predicting adolescent suicide
[57] S. Ji et al., “Knowledge transferring via model aggregation for attempts with neural networks,” in Proc. NIPS Workshop Mach. Learn.
online social care,” 2019, arXiv:1905.07665. [Online]. Available: Health, 2017, pp. 1–8.
http://arxiv.org/abs/1905.07665 [80] J. P. Pestian et al., “Sentiment analysis of suicide notes: A shared task,”
[58] M. Gaur et al., “Knowledge-aware assessment of severity of suicide Biomed. Informat. Insights, vol. 5s1, Jan. 2012, Art. no. BII.S9042.
risk for early intervention,” in Proc. World Wide Web Conf. (WWW), [81] E. White and L. J. Mazlack, “Discerning suicide notes causality using
2019, pp. 514–525. fuzzy cognitive maps,” in Proc. IEEE Int. Conf. Fuzzy Syst. (FUZZ-
[59] G. Coppersmith, R. Leary, P. Crutchley, and A. Fine, “Natural language IEEE), Jun. 2011, pp. 2940–2947.
processing of social media as screening for suicide risk,” Biomed.
[82] B. Desmet and V. Hoste, “Emotion detection in suicide notes,” Expert
Informat. Insights, vol. 10, Jan. 2018, Art. no. 117822261879286.
[60] R. Sawhney, P. Manchanda, P. Mathur, R. Shah, and R. Singh, Syst. Appl., vol. 40, no. 16, pp. 6351–6358, Nov. 2013.
“Exploring and learning suicidal ideation connotations on social media [83] R. Wicentowski and M. R. Sydes, “Emotion detection in suicide notes
with deep learning,” in Proc. 9th Workshop Comput. Approaches using maximum entropy classification,” Biomed. Informat. Insights,
Subjectivity, Sentiment Social Media Anal., 2018, pp. 167–175. vol. 5s1, Jan. 2012, Art. no. BII.S8972.
[61] M. M. Tadesse, H. Lin, B. Xu, and L. Yang, “Detection of suicide [84] A. Kovačević, A. Dehghan, J. A. Keane, and G. Nenadic, “Topic
ideation in social media forums using deep learning,” Algorithms, categorisation of statements in suicide notes with integrated rules and
vol. 13, no. 1, p. 7, Dec. 2019. machine learning,” Biomed. Informat. Insights, vol. 5s1, Jan. 2012,
[62] S. Ji, X. Li, Z. Huang, and E. Cambria, “Suicidal ideation and Art. no. BII.S8978.
mental disorder detection with attentive relation networks,” 2020, [85] A. M. Schoene and N. Dethlefs, “Automatic identification of suicide
arXiv:2004.07601. [Online]. Available: http://arxiv.org/abs/2004.07601 notes from linguistic and sentiment features,” in Proc. 10th SIGHUM
[63] A. Zirikly, P. Resnik, O. Uzuner, and K. Hollingshead, “Clpsych Workshop Lang. Technol. Cultural Heritage, Social Sci., Humanities,
2019 shared task: Predicting the degree of suicide risk in Reddit 2016, pp. 128–133.
posts,” in Proc. 6th Workshop Comput. Linguistics Clin. Psychol., 2019, [86] J. Robinson et al., “Social media and suicide prevention: A systematic
pp. 24–33. review,” Early Intervent Psychiatry, vol. 10, no. 2, pp. 103–121,
[64] A. G. Hevia, R. C. Menéndez, and D. Gayo-Avello, “Analyzing the Apr. 2016.
use of existing systems for the CLPsych 2019 shared task,” in Proc. [87] Y. Wang, S. Wan, and C. Paris, “The role of features and context on
6th Workshop Comput. Linguistics Clin. Psychol., 2019, pp. 148–151. suicide ideation detection,” in Proc. Australas. Lang. Technol. Assoc.
[65] M. Morales, P. Dey, T. Theisen, D. Belitz, and N. Chernova, Workshop, 2016, pp. 94–102.
“An investigation of deep learning systems for suicide risk assessment,” [88] A. Shepherd, C. Sanders, M. Doyle, and J. Shaw, “Using social media
in Proc. 6th Workshop Comput. Linguistics Clin. Psychol., 2019, for support and feedback by mental health service users: Thematic
pp. 177–181. analysis of a Twitter conversation,” BMC Psychiatry, vol. 15, no. 1,
[66] M. Matero et al., “Suicide risk assessment with multi-level dual-
p. 29, Dec. 2015.
context language and,” in Proc. 6th Workshop Comput. Linguistics Clin.
Psychol., 2019, pp. 39–44. [89] M. De Choudhury and S. De, “Mental health discourse on reddit: Self-
[67] L. Chen, A. Aldayel, N. Bogoychev, and T. Gong, “Similar minds post disclosure, social support, and anonymity,” in Proc. ICWSM, 2014,
alike: Assessment of suicide risk using a hybrid model,” in Proc. 6th pp. 71–80.
Workshop Comput. Linguistics Clin. Psychol., 2019, pp. 152–157. [90] M. De Choudhury, E. Kiciman, M. Dredze, G. Coppersmith, and
[68] X. Zhao, S. Lin, and Z. Huang, “Text classification of micro-blog’s M. Kumar, “Discovering shifts to suicidal ideation from mental health
‘Tree Hole’ based on convolutional neural network,” in Proc. Int. Conf. content in social media,” in Proc. CHI Conf. Hum. Factors Comput.
Algorithms, Comput. Artif. Intell. (ACAI), 2018, p. 61. Syst., May 2016, pp. 2098–2110.
[69] T. Tran, D. Phung, W. Luo, R. Harvey, M. Berk, and S. Venkatesh, [91] M. Kumar, M. Dredze, G. Coppersmith, and M. De Choudhury,
“An integrated framework for suicide risk prediction,” in Proc. 19th “Detecting changes in suicide content manifested in social media
ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining (KDD), 2013, following celebrity suicides,” in Proc. 26th ACM Conf. Hypertext Social
pp. 1410–1418. Media (HT), 2015, pp. 85–94.
[70] S. Berrouiguet et al., “Toward E-Health applications for suicide pre- [92] L. Guan, B. Hao, Q. Cheng, P. S. Yip, and T. Zhu, “Identifying chinese
vention,” in Proc. IEEE 1st Int. Conf. Connected Health, Appl., Syst. microblog users with high suicide probability using Internet-based
Eng. Technol. (CHASE), Jun. 2016, pp. 346–347. profile and linguistic features: Classification model,” JMIR Mental
[71] D. Meyer et al., “Development of a suicidal ideation detection Health, vol. 2, no. 2, p. e17, May 2015.
tool for primary healthcare settings: Using open access online psy- [93] S. J. Cash, M. Thelwall, S. N. Peck, J. Z. Ferrell, and
chosocial data,” Telemedicine e-Health, vol. 23, no. 4, pp. 273–281, J. A. Bridge, “Adolescent suicide statements on MySpace,” Cyberpsy-
Apr. 2017. chology, Behav., Social Netw., vol. 16, no. 3, pp. 166–174, Mar. 2013.

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

JI et al.: SID: A REVIEW OF MACHINE LEARNING METHODS AND APPLICATIONS 13

[94] I. Gilat, Y. Tobin, and G. Shahar, “Offering support to suicidal Xue Li (Member, IEEE) received the Ph.D. degree
individuals in an online support group,” Arch. Suicide Res., vol. 15, from the Queensland University of Technology,
no. 3, pp. 195–206, Jul. 2011. Brisbane, QLD, Australia, in 1997.
[95] A. Yates, A. Cohan, and N. Goharian, “Depression and self-harm He is currently a Professor with the School
risk assessment in online forums,” in Proc. Conf. Empirical Methods of Information Technology and Electrical Engi-
Natural Lang. Process., 2017, pp. 2968–2978. neering (ITEE), The University of Queensland,
[96] Y. Wang et al., “Understanding and discovering deliberate self-harm Brisbane. His current research interests include data
content in social media,” in Proc. 26th Int. Conf. World Wide Web, mining, social computing, database systems, and
Apr. 2017, pp. 93–102. intelligent web information systems.
[97] Q. Li, Y. Xue, L. Zhao, J. Jia, and L. Feng, “Analyzing and identifying
Teens’ stressful periods and stressor events from a microblog,” IEEE
J. Biomed. Health Informat., vol. 21, no. 5, pp. 1434–1448, Sep. 2017.
[98] Z. Huang, J. Yang, F. van Harmelen, and Q. Hu, “Constructing
knowledge graphs of depression,” in Proc. Int. Conf. Health Inf. Sci.
Cham, Switzerland: Springer, 2017, pp. 149–161.
[99] F. Hao, G. Pang, Y. Wu, Z. Pi, L. Xia, and G. Min, “Providing
appropriate social support to prevention of depression for highly Erik Cambria (Senior Member, IEEE) received
anxious sufferers,” IEEE Trans. Comput. Social Syst., vol. 6, no. 5, the Ph.D. degree from a joint program between the
pp. 879–887, Oct. 2019. University of Stirling, Stirling, U.K., and the MIT
[100] A. E. Aladaǧ, S. Muderrisoglu, N. B. Akbas, O. Zahmacioglu, and Media Lab, Cambridge, MA, USA.
H. O. Bingol, “Detecting suicidal ideation on forums: Proof-of-concept He worked at Microsoft Research Asia and
study,” J. Med. Internet Res., vol. 20, no. 6, p. e215, Jun. 2018. HP Labs India. He is the Founder of Sentic-
[101] D. N. Milne, G. Pink, B. Hachey, and R. A. Calvo, “CLPsych 2016 Net, Singapore, a Singapore-based company offering
shared task: Triaging content in online peer-support forums,” in Proc. B2B sentiment analysis services, and an Associate
3rd Workshop Comput. Lingusitics Clin. Psychol., 2016, pp. 118–127. Professor with Nanyang Technological University
[102] K. Haerian, H. Salmasian, and C. Friedman, “Methods for identifying (NTU), Singapore, where he also holds the appoint-
suicide or suicidal ideation in EHRs,” in Proc. AMIA Annu. Symp., ment of Provost Chair in Computer Science and
American Medical Informatics Association, 2012, p. 1244. Engineering.
[103] D. E. Losada and F. Crestani, “A test collection for research on Dr. Cambria is involved in many international conferences as a PC member
depression and language use,” in Proc. Int. Conf. Cross-Lang. Eval. and the program chair. He is also an Associate Editor of several journals, e.g.,
Forum Eur. Lang. Cham, Switzerland: Springer, 2016, pp. 28–39. Neurocomputing (NEUCOM), Information Fusion (INFFUS), Knowledge-
[104] G. Coppersmith, M. Dredze, C. Harman, K. Hollingshead, and Based Systems (KBS), IEEE Computational Intelligence Magazine (CIM), and
M. Mitchell, “CLPsych 2015 shared task: Depression and PTSD on the IEEE I NTELLIGENT S YSTEMS (where he manages the Department of
Twitter,” in Proc. 2nd Workshop Comput. Linguistics Clin. Psychol., Affective Computing and Sentiment Analysis).
Linguistic Signal Clin. Reality, 2015, pp. 31–39.
[105] X. Liu et al., “Proactive suicide prevention online (PSPO): Machine
identification and crisis management for chinese social media users
with suicidal thoughts and behaviors,” J. Med. Internet Res., vol. 21,
no. 5, May 2019, Art. no. e11705.

Guodong Long received the Ph.D. degree from the

University of Technology Sydney (UTS), Ultimo,
NSW, Australia, in 2014.
Shaoxiong Ji received the bachelor’s degree from He is currently a Senior Lecturer with the Aus-
the Dalian University of Technology, Dalian, China. tralian Artificial Intelligence Institute, Faculty of
He is currently pursuing the Ph.D. degree with the Engineering and IT, UTS. His research focuses on
Department of Computer Science, Aalto University, data mining, machine learning, and natural language
Espoo, Finland. processing (NLP). He has more than 40 research
He was a Research Assistant or a Visiting articles published on top-tier journals, including
Researcher with the University of Technology Syd- the IEEE T RANSACTIONS ON PATTERN A NALYSIS
ney, Ultimo, NSW, Australia, The University of AND M ACHINE I NTELLIGENCE (TPAMI), the IEEE
Queensland, Brisbane, QLD, Australia, and Nanyang T RANSACTIONS ON C YBERNETICS (TCYB), and the IEEE T RANSACTIONS
Technological University, Singapore. His research ON K NOWLEDGE AND D ATE E NGINEERING (TKDE), and conferences,
interests include machine learning and data mining. including the International Conference on Learning Representations (ICLR),
the AAAI Conference on Artificial Intelligence (AAAI), the International
Joint Conference on Artificial Intelligence (IJCAI), and the IEEE International
Conference on Data Mining (ICDM).
Shirui Pan (Member, IEEE) received the Ph.D.
degree in computer science from the University of
Technology Sydney (UTS), Ultimo, NSW, Australia,
in 2015.
He is currently a Lecturer with the Faculty of
Information Technology, Monash University, Mel- Zi Huang (Member, IEEE) received the B.Sc.
bourne, VIC, Australia. Since 2010, he has pub- degree in computer science from Tsinghua Uni-
lished over 80 research articles in top-tier journals versity, Beijing, China, and the Ph.D. degree in
and conferences, including the IEEE T RANSAC - computer science from the School of Information
TIONS ON N EURAL N ETWORKS AND L EARNING Technology and Electrical Engineering, The Univer-
S YSTEMS (TNNLS), the IEEE T RANSACTIONS ON sity of Queensland, Brisbane, QLD, Australia.
K NOWLEDGE AND D ATA E NGINEERING (TKDE), the IEEE T RANSACTIONS She is currently an ARC Future Fellow with
ON C YBERNETICS (TCYB), Pattern Recognition, the International Joint the School of Information Technology and Elec-
Conference on Artificial Intelligence (IJCAI), the International Conference trical Engineering, The University of Queensland.
on Data Engineering (ICDE), and the IEEE International Conference on Data Her research interests include multimedia indexing
Mining (ICDM). His current research interests include data mining, machine and search, social data analysis, and knowledge
learning, and graph data analytics. discovery.

Authorized licensed use limited to: UNIVERSITY OF WESTERN ONTARIO. Downloaded on October 11,2020 at 11:24:17 UTC from IEEE Xplore. Restrictions apply.