Teenagers’ Stress Detection Based

on Time-Sensitive Micro-blog
Comment/Response Actions

Liang Zhao, Jia Jia, and Ling Feng(B)

Department of Computer Science and Technology, Tsinghua University,

Beijing 100084 , China
[email protected], {jjia,fengling}@tsinghua.edu.cn

Abstract. Accurately detecting psychological stress in time is a signifi-

cant issue in the modern stressful society, especially for adolescents who
are not mature enough to cope with pressure well. Micro-blog offers a
new channel for teens’ stress detection, since more and more teenagers
nowadays prefer to express themselves on the lively virtual social net-
works. Previous work mainly rely on tweeting contents to detect tweeters’
psychological stress. However, a tweet is limited to 140 characters, which
are too short to provide enough information to accurately figure out
its tweeter’s stress. To overcome the limitation, this paper proposes to
leverage details of social interactions between tweeters and their follow-
ing friends (i.e., time-sensitive comment/response actions under a tweet)
to aid stress detection. Experimental results through a real user study
show that time sensitivity of comment/response acts plays a significant
role in stress detection, and involving such interaction acts can improve
the detection performance by 23.5% in F-measure over that without such
interactions.

1 Introduction
The increasingly faster life pace in the competitive society often makes people
stressful, especially for teenagers who are too immature to deal with psycholog-
ical pressures. Currently, 20% teenagers have psychological illness around the
world [2]. An online survey of 1018 U.S. teens (aged 13-17) made by the Amer-
ican Psychological Association in August, 2013 found that teens suffered stress
in all areas of their lives, from school to friends, work and family, which nega-
tively affected every aspect of their lives, and about 27% of the teens experienced
extreme stress and 55% experienced moderate stress in the past school year [4].
If such pressures cannot get properly relieved in time, the teenagers will suf-
fer severe physical and mental problems under accumulated pressures, such as
clinical depressions, insomnia, and even committing suicide. According to China
Center for Disease Control and Prevention [1], suicide has become the top cause
of death among Chinese youth, and excessive stress is considered to be a major
factor of suicide. Also in Korea, suicide has become teenagers’ No.1 killer in

c IFIP International Federation for Information Processing 2015
T. Dillon (Ed.): IFIP AI 2015, IFIP AICT 465, pp. 26–36, 2015.
DOI: 10.1007/978-3-319-25261-2 3
 Teenagers’ Stress Detection Based on Time-Sensitive Micro-blog 27

the past two years [3]. Annual increase of adolescent suicide rate has become a
world-wide common problem.
As adolescence is a critical period for one’s growth and development, it has
significant value to pay attention to teenagers’ psychological status and dis-
cover their suffering pressures in time. With the popularity of social networks,
micro-blog offers another low-cost sensing channel to analyze teenagers’ psycho-
logical pressures through their tweets, since more and more teenagers turn to
micro-blog for information acquisition, personal interaction, self-expression, and
emotion release. Some research work has already made efforts to detect user’s
stress and depression through micro-blog by analyzing their tweeting behav-
iors [8,9,13,15,18]. Leveraging tweeting contents such as linguistic and visual
features to analyze user’s stress has been proven feasible [11,16]. However, a
tweet faces the limitation of only 140 characters, which is too short to provide
enough content information to figure out stress and sometimes users may not
express their stress so directly. To address the limitation, we propose to further
involve comment/response acts under the tweet (comments, responses, likes and
forwards) to better supplement stress detection. Such interactions are attached
with unique timestamps. Considering time sensitivity, we select the interactions
within a certain time gap right after the tweet is posted and construct a time-
sensitive feature space for stress detection together with content of the tweet.
Based on [16], we combine the content of both the tweet and time-sensitive com-
ment/response acts under the tweet to extract content features for the four stress
categories (academic, affection, interpersonal, self-cognition), respectively. From
the observations that stressful tweets often receive more comments from friends,
especially comments with care, comfort and encourage, and the psychological
study [5] that users get inactive when suffering stress, we extract several novel
interacting behavior features such as effective comment rate, reply rate, effective
reply rate, and average interaction depth to help improve the stress detection.
Our user study of 36 high school teenagers shows that time sensitivity of com-
ment/response acts plays a significant role in stress detection and such acts in
30 minutes after the tweet posted are proved to be the most effective. Involving
such interaction acts improves the detection performance by 23.5%.
To the best of our knowledge, this is the first work combining tweet contents
and time-sensitive comment/response acts under the tweet for stress detection.
There are also no previous stress detection work, specifying features for different
stress categories.
The remainder of the paper is organized as follows. We review the related
work in Section 2. Section 3 analyzes and extracts features from time-sensitive
comment/response acts for teens’ stress detection. Experimental performance
through a real user study is evaluated in Section 4. Finally, we conclude the
paper in Section 5.

2 Related Work
Computer-aided sentiment analysis and applications in social network has drawn
much attention in recent years [10,11,16,19]. Many studies focus on the sin-
28 L. Zhao et al.

gle tweet, leveraging content features such as text-based linguistic attributes

and visual factors such as emoticons and images. [20] proposed a system called
Moodlens to do sentiment analysis for Chinese Weibo via emoticons in tweets. The
emoticons are divided into four different sentiment categories (i,e, angry, disgust-
ing, joyful and sad), and a fast Naive Bayes classifiers works for the sentiment anal-
ysis of tweets.
Beside the content features, simple social interaction attributes are also
involved. [12] proposed to detect user’s psychological stress from social media via
a deep convolution network on sequential time series in a certain time period.
Simple social features like the number of -mentions, -replies, comments and likes
are considered. User-level social connections such as (mutually) follow relation-
ships and (mutually) -mentioned relationships were also investigated in [14] to
improve sentiment analysis in Twitter. [17] explored comments to help predict
emotions expressed by images posted on Flickr. However, time sensitivity of such
social interaction attributes is not considered in these work.
The most closest related work is [11,16]. Focusing on the four main kinds
of stress (academic, affection, interpersonal and self-cognition) that troubles
teenagers, [16] extracted different features from teenagers’ tweets, such as nega-
tive emotion words, negative emoticons, unusual post time and post frequency,
etc., and several classifiers were leveraged to learn the potential stress category
and corresponding stress level (0-5) behind the tweet. However, [16] only con-
sidered the content and unusual posting behavior of the tweet, ignoring social
interactions such as comments and replies under the tweet. Besides, the detection
process only returns one stress category and one level for a single tweet, but actu-
ally a single tweet may express several kinds of stress with different levels. [11]
designed a deep sparse neural network to detect stress for arbitrary micro-blog
users, using content features of the tweet including linguistic attributes as [16],
visual attribute like color theme, brightness, etc, and simple social attributes like
number of comments, retweets and favorites. they do consider the social factors,
but ignore the details such as content of comments and the interaction of friends’
comments and author’s reply under the tweet . They only detected whether the
user suffers stress, but did not quantitatively measure the stress level. Similarly,
[11] also returned one stress category for one single tweet. In addition, for both
the two work, content features are not category-specified and features are all
static, no time-sensitive factor considered.

3 Stress Detection with Time-Sensitive

Comment/Response Acts
Before introducing the details of stress detection, we first illustrate some nota-
tions used in the following of the paper by Table 1. And without loss of generality,
we use interaction acts and comment/response acts interchangeably in the rest of
the paper. Given a tweet p = Tw (u, t0 , cont, SI Δt ), four kinds of interaction acts
under p constitute SI Δt , comments, likes, and forwards from friends, as well as
 Teenagers’ Stress Detection Based on Time-Sensitive Micro-blog 29

user’s responses. When Δt is big enough, then SI Δt contains all the comments,
replies, likes and forwards under p.
As in [16], let Category ={academic, affection,interpersonal,self-cognition} be
the set of teenagers’ stress categories involved in the study, and Level ={null,
light, moderate,strong} be the set of stress levels, where null means no-stress. For
a single tweet p, we leverage time-sensitive interaction acts SI Δt together with
content of p to form a feature space. Exploiting classification methods, the stress
detection result of p is represented as Stress(p) = ((Ci , Lv1 ), . . . , (C4 , Lv4 )),
where Ci ∈ Category and Lvi ∈ Level . If ∀Lvi = null (i = 1, . . . , 4), then Stress(p)
is a zero vector and p is a non-stressful tweet. Without loss of generality, we call
those tweets with non-zero stress vector stressful posts.

Table 1. Notations covered in tweets

Notation Representation Description

A tweet posted by user u at time t0
with the content of cont, and set of
p p = Tw (u, t0 , cont, SI Δt )
time-sensitive interactions acts SI Δt
during [t0 , t0 + Δt]
F F = {f | f is a friend of u} Friend set of user u
A comment from f at time tc under
c c = comm(f, tc , cont, p), f ∈ F tweet p, and cont is the content of
the comment
u’s reply to comment c at time tr
r r = rep(u, tr , c, cont, p) under tweet p, and cont is the
content of the reply
lk lk = like(f, tl , p), f ∈ F Friend f puts a like seal on p at tl
fw f w = f wrd(f, tf , p), f ∈ F Friend f forwards p at time tf
Comm (f, p, Δt) = {c|c.tc ∈ [t0 , t0 + Δt]} Set of comments from f within Δt
Rep (u, f, p, Δt) = {r|r.tr ∈ [t0 , t0 + Δt]} Set of replies to f within Δt
Likes (F, p, Δt) = {lk|∀f ∈ F, lk.tl ∈ [t0 , t0 + Δt]} Set of likes within Δt
Fwrds (F, p, Δt) = {f w|∀f ∈ F, f w.tf ∈ [t0 , t0 + Δt]} Set of forwards of p within Δt

3.1 Modified Content Features

A tweet is only 140 characters limited which is too short to provide enough
content information to figure out stress and sometimes users are not likely to
express their stress so directly via a brief tweet. Through daily observations,
content of the communication between the user and his/her friends under the
tweet provides powerful cues to help analyze stress. We consider the content
of user interactions with friends (i.e., comments from friends and the user’s
replies) as part of the tweet to supplement more content information. Assume
the combined content of p is denoted as
 
CCont(p, Δt) = {c.cont, r.cont|c ∈ Comm (f, p, Δt), r ∈ Rep (u, f, p, Δt)} p.cont
f ∈F

Particularly, interactive comments and responses between user u and friend f

are closely related in content and constitute a dialog. We take the content of
30 L. Zhao et al.

dialog as a whole, and regard it as an item piece of CCont(p, Δt). For each item
piece in CCont(p, Δt), we leverage a graph-based Chinese parser [6,7] to analyze
linguistic associations, including the correspondence between stress categories
and negative emotion words, and the grammatical relation between adverbs of
degree and the modified negative emotion words, etc. A single tweet may con-
vey several kinds of stress at the same time. Thus, for each stress category in
Category , we specify content features from CCont(p, Δt) as in [16]. i.e., number
of negative emotion words, number of positive and negative emoticons, number
of exclamation and question marks, emotional degree, etc.

3.2 Time-Sensitive Comment/Response Features

Comment/Response acts under a tweet is time-sensitive especially when the

user is suffering from stress. For example, when a user posts a tweet expressing
depression, his/her friends will always make in-time comments within a short
time gap after the tweet is released. We select the interaction acts under the
tweet p within [t0 , t0 + Δt] to extract the comment/response features.

Commenting Features from Crowd. Comments, likes and forwards are basic
commenting features and always show the social attention to a tweet.
Number of Comments from Friends. The set of all the comments under the
tweet p within [t0 , t0 + Δt] is

Comments (F, p, Δt) = Comm (f, p, Δt)
f ∈F

Number of Likes. People always put a like seal on a tweet to show their
positive emotion or attitude to the tweet. Compared with those positive tweets,
stressful tweets (negative tweets) obtain much less likes, even no like. We take
the number of likes under a tweet as an indicator of potential stress, denoted as
|Likes (F, p, Δt)|.
Number of Forwards. Users tend to forward public information such as news,
advertisements and jokes, and seldom retweet personalized stressful tweets of
another user. So stressful tweets enjoy less possibility to be forwarded by friends.
The number of forwards here is denoted as |Fwrds (F, u, p, Δt)|.
Number of Effective Comments. Among all the comments from friends,
we further check the details of the comments and distinguish those contain-
ing care, comfort, or encourage (e.g., “What’s up?”, “Don’t worry”, “Every-
thing will be OK”, or hug emoticon, etc.). Such comments are considered as
effective comments. The set of effective comments under tweet p is denoted as
EComments (F, p, Δt). A lexicon with 132 such Chinese phrases, sentences and
emoticons is constructed to identify effective comments. A tweet with more effec-
tive comments is more likely to be stressful.
 Teenagers’ Stress Detection Based on Time-Sensitive Micro-blog 31

Effective Comment Rate.(short for ECR, denoted as ECR(F, p, Δt)) is cal-

culated as
|EComments (F, p, Δt)|
ECR(F, p, Δt) =
|Comments (F, p, Δt)| + 1
where |(·)| stands for the number of items in set (·) and we smooth the equa-
tion by adding 1 upon denominator. Obviously, a higher ECR denotes a bigger
probability that the tweet is stressful.
Particularly, with any two of the three features (number of com-
ments/effective comments and effective comment rate) we can deduce the other
one. Thus, for commenting features we can choose any two.

Replying Features of the User. Besides the social behaviors of the crowd
(friends), behaviors of the user himself also make sense in stress detection. Psy-
chological study [5] shows that a user in stress may perform low activeness
in social networks. Such inactiveness can be revealed from user’s reactions to
the comments from friends. Three features corresponding to the user’s replying
behaviors are identified.
Reply Rate. If a user is absorbed in big stress, s/he tends to be inactive, which
can be revealed from his/her activeness of reply to the comments. The reply rate
(short for RR, denoted as RR(u, p, Δt)) is calculated by the rate of replies over
total number of comments. Assume Replies (u, F, p, Δt) = f ∈F Rep (u, f, p, Δt)
is the set of user replies to all the friends under p, then
|Replies (u, F, p, Δt)|
RR(u, p, Δt) =
|Comments (F, p, Δt)| + 1
A lower reply rate suggests that the user is not so active and it is highly
possible that s/he is suffering stress. Similarly, among the three features (number
of comments/replies, and reply rate), we can deduce the other one with any two.
Effective Reply Rate. Effective replies are those responses to the effective
comments. The set of effective replies under p within [t0 , t0 + Δt] is
EReplies (u, F, p, Δt) = {r|r.c ∈ EComments (F, p, Δt), r.tr ∈ [t0 , t0 + Δt]}

We measure the effective reply rate (short for ERR, denoted as ERR(p, Δt)) by
the proportion of effective reply over the total number of effective comments.
|EReplies (u, F, p, Δt)|
ERR(p, Δt) =
|EComments (F, p, Δt)| + 1

Average Interaction Depth with Friends. A friend may comment multiple

times under the same tweet, and the user may also make corresponding replies.
Then, such interaction between them constitutes a dialog. In the dialog, once
a new comment-reply pair comes, the interaction depth between the friend and
the user accordingly increases by 1. Assume a friend f and the user u make a
32 L. Zhao et al.

comment-reply dialog under tweet p, then the interaction depth between f and
u in tweet p within [t0 , t0 + Δt] is calculated as
iDep(f, u, p, Δt) = min(|Comm (f, p, Δt)|, |Rep (u, f, p, Δt)|)
For all the friends who comment the tweet, the average interaction depth is
computed by
1 
avg iDep(F, u, p, Δt) = iDep(f, u, p, Δt)
|F |
f ∈F

The average interaction depth reveals the user’s activeness as well. The smaller
the average interaction depth is, the less active the user performs, so that s/he
has a higher risk of stress suffered.

4 Experimental Study
4.1 Setup
Psychological stress detection is a highly personalized issue and there are no
available benchmark specially for the problem. In this work, we collect data
ourselves and conduct a user study on a real micro-blog data set. 36 high school
students (15 males and 21 females, aged between 15 and 17) in Shaanxi Province,
China, participated in the user study. 21,648 tweets from 2013/1/1 to 2015/5/1
were collected from their accounts in Tencent Weibo1 , averagely 636 tweets per
teenager. Recalling the real situations, they were asked to scan their own tweets
one by one and annotated the psychological stress on the four stress categories in
Category and corresponding stress levels in Level expressed in each single tweet.
We take the annotation as the ground truth. For each teenager, chronologically,
we use the early 66% of his/her tweets as the training data, and the rest 34% as
the testing data.

4.2 General Performance of Comment/Response Acts

Four different classifiers, including Naive Bayes, Logistic, SVM and Gaussian
Process are used to perform the single-tweet based stress detection over the
feature space of each stress category, respectively. Precision and recall are lever-
aged to evaluate the performance. With the comment/response acts selected
within Δt after the tweet posted, assume for each stress category Ci ∈ Category ,
T P (Ci , Lvi , Δt), T N (Ci , Lvi , Δt), F P (Ci , Lvi , Δt), F N (Ci , Lvi , Δt) represents
the number of true positive, true negative, false positive and false negative sam-
ples detected of stress level Lvi , respectively. Thus,
1 
P recision(Ci , Δt) = precision(Ci , Lvj , Δt)
|Level |
Lvj ∈Level
1 
Recall(Ci , Δt) = recall(Ci , Lvj , Δt)
|Level |
Lvj ∈Level

1
One of the biggest Chinese micro-blog platform, http://t.qq.com/
 Teenagers’ Stress Detection Based on Time-Sensitive Micro-blog 33

T P (Ci ,Lvj ,Δt)

where precision(Ci , Lvj , Δt) = T P (Ci ,Lvj ,Δt)+F P (Ci ,Lvi ,Δt)
, and recall(Ci , Lvj , Δt)
T P (Ci ,Lvj ,Δt)
= T P (Ci ,Lvj ,Δt)+F N (Ci ,Lvj ,Δt) .

Table 2. General Performance (Δt = 30 minutes)

Naive Bayes Logistic SVM Gaussian

Stress
Prec. Rec. F-ms. Prec. Rec. F-ms. Prec. Rec. F-ms. Prec. Rec. F-ms.
Aca 0.69 0.69 0.69 0.71 0.67 0.69 0.73 0.67 0.70 0.77 0.70 0.73
Aff 0.53 0.72 0.61 0.81 0.72 0.76 0.87 0.81 0.83 0.44 0.38 0.41
Inter 0.67 0.70 0.68 0.81 0.72 0.76 0.73 0.63 0.67 0.56 0.36 0.44
Self 0.60 0.66 0.53 0.69 0.67 0.68 0.66 0.64 0.65 0.32 0.33 0.33
Avg. 0.62 0.69 0.65 0.76 0.70 0.72 0.75 0.69 0.71 0.52 0.44 0.48

Table 2 compares the performance with Δt = 30 minutes. Averagely speak-

ing, Logistic and SVM work the best in the four classifiers for all the stress
categories, with the average F-measure over 70%, which is 10.7% better than
NB, and 50% better than Gaussian.

4.3 Investigation of Time Sensitivity

We investigate the time sensitivity of comment/response acts with different Δt

value. Logistic is selected as the classifier in this experiment since it performs
the best of the three. Different time gap Δt (= 10, 20, 30, 40, and +∞ minutes)
are selected to extract comment/response features, where all the interaction acts
under a tweet are involved when Δt = +∞. Fig. 1 shows the performance of
different time sensitivity (Δt). Obviously, for each stress category, the detection
results are time-sensitive. With the increase of Δt, the performance first improves
and then falls again as Δt getting bigger. Particularly, when Δt = 30min, the
detection performance reaches a peak, with the average F-measure over 70%. The
result coincides with our daily observation. When a user posts a stressful tweet,
his/her friends often comments in time to express care, comfort and encourage.
With Δt increases within [0, 30] (counted in minute), such effective information
accumulates and makes the detection more accurate.

4.4 Impact of Comment/Response Acts

Table 3 compares the detection performance with/without comment/response

acts under Logistic classifier. We set Δt = 30 minutes for time-sensitive inter-
action acts. Due to user’s indirect expression, without interaction acts cannot
provide enough information and performs worse. When involving the comment-
/response acts, the average F-measure for all the stress categories reaches over
70%, outperforms the other case 23.5%, which proves the significance of time-
sensitive comment/response acts in stress detection.
34 L. Zhao et al.

(a) Academic Stress (b) Affection Stress

(c) Interpersonal Stress (d) Self-cognition Stress

Fig. 1. Performance with different time sensitivity (Δt)

Table 3. Comparison with/without comment/response acts

Without C/R acts With C/R acts

Stress
Prec. Rec. F-meas. Prec. Rec. F-meas.
academic 0.596 0.553 0.574 0.713 0.669 0.690
affection 0.558 0.579 0.568 0.808 0.722 0.763
interpersonal 0.744 0.613 0.669 0.808 0.722 0.763
self-cognition 0.553 0.5 0.568 0.693 0.667 0.679
avg. 0.613 0.561 0.586 0.756 0.695 0.724

5 Conclusion
In this paper, together with content of tweets, we involve time-sensitive com-
ment/response acts under a tweet to help better detect stress revealed from a
teenager’s tweet. Focusing on the four main categories of adolescent stress (aca-
demic, affection, interpersonal, and self-cognition), we specify content features
for each stress category, and involve several novel features extracted from time-
sensitive comment/response acts. 33 high school students aged 17 participated in
our user study to evaluate the framework. Experimental results show that stress
detection performance varies with different time sensitivity, and involving com-
 Teenagers’ Stress Detection Based on Time-Sensitive Micro-blog 35

ment/response acts can better improve the F-measure of multi-category stress

detection by 23.5%.

Acknowledgement. The work is supported by National Natural Science Foundation

of China (61373022, 61370023, 61073004), and Chinese Major State Basic Research
Development 973 Program (2011CB302203-2).

References
1. The week news. http://theweek.com/article/index/252199/the-rise-of-youth-suici-
de-in-china
2. Sohu news (2012). http://learning.sohu.com/s2012/shoot/
3. Sohu news (2013). http://learning.sohu.com/20130402/n371458123.shtml
4. USA today (2014). http://www.usatoday.com/story/news/nation/2014/02/11/
stress-teens-psychological/5266739
5. Social media and the cost of caring (2015). http://www.pewinternet.org/files/
2015/01/PI Social-media-and-stress 0115151.pdf
6. Che, W., Li, Z., Guo, Y., Qin, B., Liu, T.: Multilingual dependency-based syntactic
and semantic parsing. In: Proc. of CoNLL, pp. 49–54 (2009)
7. Che, W., Li, Z., Liu, T.: LTP: a chinese language technology platform. In: Proc.
of Coling, pp. 13–16 (2010)
8. Choudhury, M., Counts, S., Horvitz, E.: Social media as a measurement tool of
depression in populations. In: Proc. of ACM Web Science, pp. 47–56 (2013)
9. Choudhury, M., Gamon, M., Counts, S., Horvitz, E.: Prediction depression via
social media. In: Proc. of ICWSM (2013)
10. Li, Q., Xue, Y., Jia, J., Feng, L.: Helping teenagers relieve psychological pressures:
a micro-blog based system. In: Proc. of EDBT, pp. 660–663 (2014)
11. Lin, H., Jia, J., Guo, Q., Xue, Y., Huang, J., Cai, L., Feng, L.: Psychological stress
detection from cross-media microblog data using deep sparse neural nework. In:
Proc. of ICME, pp. 1–6 (2014)
12. Lin, H., Jia, J., Guo, Q., Xue, Y., Li, Q., Huang, J., Cai, L., Feng, L.: User-level
psychological stress detection from social media using deep neural network. In:
Proc. of MM, pp. 507–516 (2014)
13. Park, M., McDonald, D., Cha, M.: Perception differences between the depressed
and non-depressed users in twitter. In: Proc. of ICWSM (2013)
14. Tan, C., Lee, L., Tang, J., Jiang, L., Zhou, M., Li, P.: User-level sentiment analysis
incorporating social networks. In: Proc. of SIGKDD, pp. 1397–1405 (2011)
15. Wang, X., Zhang, C., Ji, Y., Sun, L., Wu, L.: A depression detection model based
on sentiment analysis in micro-blog social network. In: Proc. of PAKDD (2013)
16. Xue, Y., Li, Q., Jin, L., Feng, L., Clifton, D.A., Clifford, G.D.: Detecting adolescent
psychological pressures from micro-blog. In: Zhang, Y., Yao, G., He, J., Wang, L.,
Smalheiser, N.R., Yin, X. (eds.) HIS 2014. LNCS, vol. 8423, pp. 83–94. Springer,
Heidelberg (2014)
17. Yang, Y., Jia, J., Zhang, S., Wu, B., Chen, Q., Li, J., Xing, C., Tang, J.: How do
your friends on social media disclose your emotions. In: Proc. of AAAI
36 L. Zhao et al.

18. Shen, Y., Kuo, T., Yeh, I., Chen, T., Lin, S.: Exploiting temporal information
in a two-stage classification framework for content-based depression detection. In:
Proc. of PAKDD, pp. 276–288 (2013)
19. Zhang, Y., Tang, J., Sun, J., Chen, Y., Rao, J.: Moodcast: emotion prediction via
dynamic continuous factor graph model. In: Proc. of ICDM
20. Zhao, J., Dong, L., Wu, J., Xu, K.: Moodlens: an emoticon-based sentiment analysis
system for chinese tweets. In: Proc. of SIGKDD, pp. 1528–1531 (2012)