0% found this document useful (0 votes)

49 views

Detecting Stress Based On Social Interactions in Social Networks

This document discusses detecting stress based on social interactions in social networks. It proposes using a hybrid model combining a factor graph model and convolutional neural network to leverage content from tweets and social interaction information. The model improves stress detection performance by 6-9% in F1-score. Analysis of social interaction data found stressed users have social structures with sparser connections, indicating their friends are less connected than non-stressed users. Traditional stress detection is limited but social media data from large-scale networks provides new opportunities to model stress through online behavior and social information.

Uploaded by

shital

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

49 views

Detecting Stress Based On Social Interactions in Social Networks

Uploaded by

shital

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 14

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2017.2686382, IEEE
Transactions on Knowledge and Data Engineering
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 1

Detecting Stress Based on Social Interactions in

Social Networks
Huijie Lin, Jia Jia, Jiezhon Qiu, Yongfeng Zhang, Lexing Xie, Jie Tang, Ling Feng, and Tat-Seng Chua

Abstract—Psychological stress is threatening people’s health. It is non-trivial to detect stress timely for proactive care. With the
popularity of social media, people are used to sharing their daily activities and interacting with friends on social media platforms,
making it feasible to leverage online social network data for stress detection. In this paper, we find that users stress state is closely
related to that of his/her friends in social media, and we employ a large-scale dataset from real-world social platforms to systematically
study the correlation of users’ stress states and social interactions. We first define a set of stress-related textual, visual, and social
attributes from various aspects, and then propose a novel hybrid model - a factor graph model combined with Convolutional Neural
Network to leverage tweet content and social interaction information for stress detection. Experimental results show that the proposed
model can improve the detection performance by 6-9% in F1-score. By further analyzing the social interaction data, we also discover
several intriguing phenomena, i.e. the number of social structures of sparse connections (i.e. with no delta connections) of stressed
users is around 14% higher than that of non-stressed users, indicating that the social structure of stressed users’ friends tend to be
less connected and less complicated than that of non-stressed users.

Index Terms—Stress detection, factor graph model, micro-blog, social media, healthcare, social interaction.

1 I NTRODUCTION YvUJ.N[KQI Do~3.Z1?I

"# $%'$! ( %%!!! % ) (% !

1.1 Motivation
# #"*%!$#HH.YCXG- # #"*%!$#=aWJ.A>5F7b{j
#%&# %!&$* (!#$ KL KL
Psychological stress is becoming a threat to people’s $# #"*%!# m]_phl:P.XPRd`96+
Don’t worry, you will get better with more efforts!

$# #"*%!# \uqTV.Z1xO0N

health nowadays. With the rapid pace of life, more and $ $%' , !%!$*!&%#&$* (!# KL Feeling stupid, I had tried my best but couldn’t do well. KL

# @[email protected] #

O2^04r|<-

more people are feeling stressed. According to a worldwide #&"#(% !#*!&%! KL Scores are not everything! KL

survey reported by Newbusiness in 20101 , over half of the

$# #"*%!# cI.zkP0^3B>NM $# #"*%!#
f^RfNS8RfyRse3-
$. $%#$$ !&%%& $"#!% KL No scores, no good jobs, no money, and I will starve to death! KL

population have experienced an appreciable rise in stress # c0c;ED@iI/

*!&(!#!'#% %!" *
KL
#
#"*%! $# =w1}t-
Don’t be that Pessimistic
KL

over the last two years. Though stress itself is non-clinical

and common in our life, excessive and chronic stress can (a) (b)
be rather harmful to people’s physical and mental health.
Fig. 1. Sample tweets from Sina Weibo. In each tweet, the top part
According to existing research works, long-term stress has is tweet content with text and an image; the bottom part shows the
been found to be related to many diseases, e.g., clinical social interactions of tweets where there are multiple indicators of stress:
depressions, insomnia etc.. Moreover, according to Chinese mentions of ’busy ’ and ’stressed’, ’working overtime’, ’failed the exam’,
’money ’ and a stressed emoticon.
Center for Disease Control and Prevention, suicide has
become the top cause of death among Chinese youth, and
more and more people are willing to share their daily events
excessive stress is considered to be a major factor of suicide.
and moods, and interact with friends through the social
All these reveal that the rapid increase of stress has become
networks. As these social media data timely reflect users’
a great challenge to human health and life quality.
real-life states and emotions in a timely manner, it offers
Thus, there is significant importance to detect stress be-
new opportunities for representing, measuring, modeling,
fore it turns into severe problems. Traditional psychological
and mining users behavior patterns through the large-scale
stress detection is mainly based on face-to face interviews,
social networks, and such social information can find its
self-report questionnaires or wearable sensors. However,
theoretical basis in psychology research. For example, [7]
traditional methods are actually reactive, which are usually
found that stressed users are more likely to be socially less
labor-consuming, time-costing and hysteretic. Are there any
active, and more recently, there have been research efforts
timely and proactive methods for stress detection?
on harnessing social media data for developing mental
The rise of social media is changing people’s life, as
and physical healthcare tools. For example, [27] proposed
well as research in healthcare and wellness. With the de-
to leverage Twitter data for real-time disease surveillance;
velopment of social networks like Twitter and Sina Weibo2 ,
while [35] tried to bridge the vocabulary gaps between
health seekers and providers using the community gener-
• Huijie Lin is with the Department of Computer Science and Technology,
Tsinghua University, Beijing, China. ated health data. There are also some research works [28],
E-mail: [email protected] [47] using user tweeting contents on social media platforms
1
http://tinyurl.com/htunr9g
to detect users’ psychological stress. Existing works [28],
2
http://www.weibo.com, one of the most popular social media [47] demonstrated that leverage social media for healthcare,
platforms in China and in particular stress detection, is feasible.

thus less complicated than those of non-stressed users.

1.2 Our Work

Inspired by psychological theories, we first define a set of at-
tributes for stress detection from tweet-level and user-level
aspects respectively: 1) tweet-level attributes from content
of user’s single tweet, and 2) user-level attributes from
user’s weekly tweets. The tweet-level attributes are mainly

composed of linguistic, visual, and social attention (i.e.,
being liked, retweeted, or commented) attributes extracted
Fig. 2. The sampling test results of the diversity of users’ social struc-
tures from Sina Weibo, by using the top 3 interacted friends of the users. from a single-tweet’s text, image, and attention list. The
user-level attributes however are composed of: (a) posting
Limitations exist in tweeting content based stress de- behavior attributes as summarized from a user’s weekly tweet
tection. Firstly, tweets are limited to a maximum of 140 char- postings; and (b) social interaction attributes extracted from
acters on social platforms like Twitter and Sina Weibo, and a user’s social interactions with friends. In particular, the
users do not always express their stressful states directly in social interaction attributes can further be broken into: (i) social
tweets. Secondly, users with high psychological stress may interaction content attributes extracted from the content of
exhibit low activeness on social networks, as reported by a users’ social interactions with friends; and (ii) social interac-
recent study in Pew Research Center3 . These phenomena tion structure attributes extracted from the structures of users’
incur the inherent data sparsity and ambiguity problem, social interactions with friends.
which may hurt the performance of tweeting content based To maximally leverage the user-level information as well
stress detection performance. For illustration, let’s see a Sina as tweet-level content information, we propose a novel
Weibo tweet example in Figure 1. The tweet contains only 13 hybrid model of factor graph model combined with a con-
characters, saying that the user wished to go home for the volutional neural network (CNN). This is because CNN is
Spring Festival holiday. Although no stress is revealed from capable of learning unified latent features from multiple
the tweet itself, from the follow-up interactive comments modalities, and factor graph model is good at modeling
made by the user and her friends, we can find that the user the correlations. The overall steps are as follows: 1) we first
is actually stressed from work. Thus, simply relying on a design a convolutional neural network (CNN) with cross
user’s tweeting content for stress detection is insufficient. autoencoders (CAE) to generate user-level content attributes
Users’ social interactions on social networks contain from tweet-level attributes; and 2) we define a partially-
useful cues for stress detection. Social psychological stud- labeled factor graph (PFG) to combine user-level social inter-
ies have made two interesting observations. The first is mood action attributes, user-level posting behavior attributes and
contagion [37]: a bad mood can be transferred from one the learnt user-level content attributes for stress detection.
person to another during social interaction. The second is We evaluate the proposed model as well as the contribu-
linguistic echoes [34]: people are known to mimic the style tions of different attributes on a real-world dataset from Sina
and affect of another person. These observations motivate Weibo. Experimental results show that by exploiting the
us to expand the scope of tweet-wise investigation by incor- users’ social interaction attributes, the proposed model can
porating follow-up social interactions like comments and improve the detection performance (F1-score) by 6-9% over
retweeting activities in user’s stress detection. This may that of the state-of-art methods. This indicates that the pro-
actually help to mitigate the single user’s data sparsity posed attributes can serve as good cues in tackling the data
problem. Another reason for considering social interactions sparsity and ambiguity problem. Moreover, the proposed
in stress detection is based on our empirical findings on a model can also efficiently combine tweet content and social
large-scale dataset crawled from Sina Weibo that the social interaction to enhance the stress detection performance.
structures of stressed users are less connected and thus less We further conduct in-depth studies on a large-scale
complicated than those of non-stressed users. This is con- dataset from Sina Weibo. Beyond user’s tweeting contents,
sistent with the Pew Research Center’s finding that stressed we analyze the correlation of users’ stress states and their
users are less active than non-stressed ones. The bottom of social interactions on the networks, and address the problem
Figure 2 illustrates four social interaction structure patterns. from the standpoints of: (1) social interaction content,
Each node in a structure pattern represents a user’s interact- by investigating the content differences between stressed
ing friend (who either commented or retweeted the tweets). and non-stressed users’ social interactions; and (2) social
If two nodes are also friends on social network, there is an interaction structure, by investigating the structure differ-
edge linking both; otherwise, there is none. We examined ences in terms of structural diversity, social influence, and
3000 users on Sina Weibo. For each user, we collected and strong/weak tie. Our investigation unveils some intriguing
merged his/her one week tweets into one and sense stress social phenomena. For example, we find that the number
from it. Meanwhile, we captured the top-3 most active of social structures of sparse connection (i.e. with no delta
friends the user interacted with. As shown in Figure 2, connections4 ) of stressed users is around 14% higher than
stressed users’ interaction structures are less connected, and that of non-stressed users, indicating that the social structure
of stressed users’ friends tend to be less connected and
3
Social Media and the Cost of Caring, 2015, complicated, compared to that of non-stressed users.
http://www.pewinternet.org/files/2015/01/PI Social-media-and-
4
stress 0115151.pdf Meaning that three points are connected with each other

The contributions of this paper are as following. states from social media by learning user-level presentation
via a deep convolution network on sequential tweet series
• We propose a uniﬁed hybrid model integrating CNN
in a certain time period. Motivated by the principle of
with FGM to leverage both tweet content attributes
homophily, [38] incorporated social relationships to improve
and social interactions to enhance stress detection.
user-level sentiment analysis in Twitter. Though some user-
• We build several stressed-twitter-posting datasets by
level emotion detection studies have been done, the role that
different ground-truth labeling methods from several
social relationships plays in one’s psychological stress states, and
popular social media platforms and thoroughly eval-
how we can incorporate such information into stress detection
uate our proposed method on multiple aspects.
have not been examined yet.
• We carry out in-depth studies on a real-world large-
Research on leveraging social interactions for social
scale dataset and gain insights on correlations be-
media analysis. Social interaction is one of the most im-
tween social interactions and stress, as well as social
portant features of social media platforms. Now many
structures of stressed users.
researchers are focusing on leveraging social interaction
The rest of this paper is organized as follows. Section information to help improve the effectiveness of social me-
2 gives an overview of related works. Section 3 presents dia analysis. [12] analyzed the relationships between social
our problem statement. Then in Section 4, we introduce interactions and users’ thinking and behaviors, and found
the deﬁnitions of the proposed attributes. Section 5 presents out that Twitter-based interaction can trigger effectual cog-
the hybrid model and training method for stress detection. nitions. [49] leveraged comments on Flickr to help predict
Experimental results are shown in Section 6. Then in Section emotions expressed by images posted on Flickr. However,
7, we present several in-depth studies on our dataset for these work mainly focused on the content of social inter-
further insights. Finally, we make some conclusions and actions, e.g., textual comment content, while ignoring the
discuss in Section 8. inherent structural information like how users are connected.

2 R ELATED W ORK 3 P ROBLEM F ORMULATION

Psychological stress detection is related to the topics of Before presenting our problem statement, let’s first define
sentiment analysis and emotion detection. some necessary notations.
Research on tweet-level emotion detection in social Let V be a set of users on a social network, and let |V |
networks. Computer-aided detection, analysis, and applica- denote the total number of users. Each user vi ∈ V posts a
tion of emotion, especially in social networks, have drawn series of tweets, with each tweet containing text, image, or
much attention in recent years [8], [9], [28], [41], [52], [53]. video content; the series of tweets contribute to users social
Relationships between psychological stress and personality interactions on the social network.
traits can be an interesting issue to consider [11], [16], [43].
Definition 1. Stress state. The stress state y of user vi ∈ V at
For example, [1] providing evidence that daily stress can be
time t is represented as a triple (y, vi , t), or briefly yit . In the
reliably recognized based on behavioral metrics from users
study, a binary stress state yit ∈ {0, 1} is considered, where
mobile phone activity. Many studies on social media based
yit = 1 indicates that user vi is stressed at time t, and yit = 0
emotion analysis are at the tweet level, using text-based
indicates that the user is non-stressed at time t, which can be
linguistic features and classic classification approaches. [53]
identified from specific expressions in user tweets or clearly
proposed a system called MoodLens to perform emotion
identified by user himself, as explained in the experiments.
analysis on the Chinese micro-blog platform Weibo, clas-
Let Y t be the set of stress states of all users at time t.
sifying the emotion categories into four types, i.e., angry,
disgusting, joyful, and sad. [9] studied the emotion propaga- Definition 2. Time-varying user-level attribute matrix.
tion problem in social networks, and found that anger has a Each user in V is associated with a set of attributes A. Let
stronger correlation among different users than joy, indicat- X t be a |V | × |A| attribute matrix at time t, in which every
ing that negative emotions could spread more quickly and row xti corresponds to a user, each column corresponds to
broadly in the network. As stress is mostly considered as a an attribute, and an element xti,j is the j -th attribute value
negative emotion, this conclusion can help us in combining of user vi at time t.
the social influence of users for stress detection. However,
A user-level attribute matrix describes user-specific featu
these work mainly leverage the textual contents in social
res, and can be defined in different ways. This study con-
networks. In reality, data in social networks is usually com-
siders user-level content attributes, statistical attributes, and
posed of sequential and inter-connected items from diverse
social interaction attributes. A detailed discussion of the
sources and modalities, making it be actually cross-media
matrix can be found in Section 4.
data.
Research on user-level emotion detection in social Definition 3. Time-varying edge set. Users are linked by
networks. While tweet-level emotion detection reflects the edges of certain types. Let E t ⊆ V × V × C be a set of edges
instant emotion expressed in a single tweet, people’s emo- between users at time t. Three types of edges are considered
tion or psychological stress states are usually more endur- in the study. For an edge e = (vi , vj , c) ∈ E t , c = 0 indicates
ing, changing over different time periods. In recent years, that vi follows or is followed by vj at time t, c = 1 indicates
extensive research starts to focus on user-level emotion that there are positive words in comments between user vi
detection in social networks [29], [36], [38], [50]. Our recent and vj at time t, and c = 2 indicates that there are negative
work [29] proposed to detect users psychological stress words in comments between them at time t.

TABLE 1
Summary of tweet-level attributes. The column “#“ indicates the feature vector length for each type of feature

Category Short Name # Description

Positive & Negative Emotion Words 2 Number of positive and negative emotion words
Positive & Negative Emoticons 2 Number of popular positive and negative emoticons, e.g., and
Linguistic
To signify the intensity of emotion four typical punctuation marks
Punctuation Marks & Associated Emotion Words 4
(’!’, ’?’, ’...’, ’.’) are considered.
In examples “{I feel a little bit sad}” and “{I feel terribly sad}” , ’sad’
expresses different negative feelings. We use 1-3 to represent neutral,
Degree Adverbs & Associated Emotion Words 2
moderate, and severe degree of positive emotions, and the minus to
represent the negative ones.
A combination of ﬁve dominant colors in HSV color space, indicating
Five-color theme 15 main color distribution of images, has been revealed to be
important on human emotions by psychology and art theories.
Visual Saturation 2 The mean value of saturation and its contrast.
Brightness 2 The mean value of brightness and its contrast.
Warm / Cool color 1 Ratio of cool colors with hue ([0-360]) in the HSV space in [30, 110].
Clear / Dull color 1 Ratio of colors with brightness ([0-1]) and saturation < 0.6.
Social Social Attention 3 Number of comments, retweets, and likes

Definition 4. Time-varying attribute-augmented network. degree adverbs. We have also tested other linguistic re-
An attribute-augmented network at time t is comprised of sources including NRC5 and HowNet6 , and found that the
four elements, including 1) a user set V t , 2) an edge set E t , performances were relatively the same, so we adopted the
3) a user-level attribute matrix set X t , and 4) a stress state set commonly used LIWC2007 dictionary for experiments. Fur-
for all users Y t at time t, denoted as Gt = (V t , E t , X t , Y t ). thermore, we extract linguistic attributes of emoticons (e.g.,
and ) and punctuation marks (’!’, ’?’, ’...’, ’.’). Weibo
Given a sequence of labeled time-varying attribute-
defines every emoticon in square brackets (e.g., they use
augmented networks at different times, our goal is to learn a
[haha] for “laugh”), so we can map the keyword in square
model that can best fit the relationships among users’ stress
brackets to find the emoticons. Twitter adopts Unicode
states, user-level attributes, and users’ social linkage, and
as the representation for all emojis [15], [24], which can
then detect users’ unknown stress states with the model.
be extracted directly. The list of linguistic attributes and
Problem 1. Psychological stress detection: Given a series descriptions are shown in Table 1.
of T partially labeled time-varying attribute-augmented net- As for the visual attributes, we use API from OpenCV 7
works {Gt = (VLt , VUt , E t , YLt ) | t ∈ {1, 2, · · · , T }}, VLt is a to perform picture processing and color-related attributes
set of users with labeled stress states YLt at time t, and VUt is computation, e.g., saturation, brightness, warm/cool color,
a set of unlabeled users, the objective is to learn a function clear/dull color in Table 1. For a special class of attributes
named five-color theme, we adopt algorithm from papers
f : {G1 , G2 , . . . GT } → {YU1 , YU2 , . . . YUT } on affective image classification [32] and color psychology
theories [23], [45]. In this work, we did not adopt the
to predict unlabeled users’ stress states.
direct emotional detection results as visual features because
we need multi-dimensional visual features for deep model
learning, while a direct visual emotional classification result
4 ATTRIBUTES C ATEGORIZATION AND D EFINITION only gives a single or very few dimensions as features.
To address the problem of stress detection, we first define However, with the development of emotion-sensitive visual
two sets of attributes to measure the differences of the representation techniques, it would be possibility to adopt
stressed and non-stressed users on social media platforms: automatic visual features in the future. The details of tweet-
1) tweet-level attributes from a user’s single tweet; 2) user- level attributes are summarized in Table 1.
level attributes summarized from a user’s weekly tweets.
4.2 User-Level Attributes
4.1 Tweet-level Attributes Compared to tweet-level attributes extracted from a sin-
Tweet-level attributes describe the linguistic and visual gle tweet, user-level attributes are extracted from a list of
content, as well as social attention factors (being liked, user’s tweets in a specific sampling period. We use one
commented, and retweeted) of a single tweet. week as the sampling period in this paper. On one hand,
For linguistic attributes, we take the most commonly psychological stress often results from cumulative events or
used linguistic features in sentiment analysis research. mental states. On the other hand, users may express their
Specifically, we first adopt LTP [4] — A Chinese Language chronic stress in a series of tweets rather than one. Besides,
Technology Platform — to perform lexical analysis, e.g., the aforementioned social interaction patterns of users in
tokenize and lemmatize, and then explore the use of a a period of time also contain useful information for stress
Chinese LIWC dictionary — LIWC2007 [14], to map the detection. Moreover, as aforementioned, the information in
words into positive/negative emotions. LIWC2007 is a dic- 5
http://www.saifmohammad.com/WebPages/NRC-Emotion-
tionary which categorizes words based on their linguistic Lexicon.htm
or psychological meanings, so we can classify words into 6
http://www.keenage.com
different categories, e.g. positive/negative emotion words, 7
http://opencv.org

tweets is limited and sparse, we need to integrate more The user-level attributes of a user at time t are denoted
complementary information around tweets, e.g., users’ so- by xti (i=1,2,· · · ) in Figure 3. The route of the other users’
cial interactions with friends. attributes in Figure 3 are similar, which finally form their
Thus, appropriately designed user-level attributes can user-level attributes. We focus on the attribute flow of the
provide a macro-scope of a user’s stress states, and avoid user with red star and omit the detailed route of other users’
noise or missing data. Here, we define user-level attributes attributes in the figure. The stress state of each user at time
from two aspects to measure the differences between t is denoted by yit (i=1,2,· · · ), respectively. The user-level
stressed and non-stressed states based on users’ weekly attributes and the stress states are connected by an attribute
tweet postings: 1) user-level posting behavior attributes [29] factor, while stress states of different users are connected by
from the user’s weekly tweet postings; and 2) user-level social factors. Stress states of the same user at adjacent times
social interaction attributes from the user’s social interactions are connected by dynamic factors. We define the graph as a
beneath his/her weekly tweet postings. The details of user- (PFG). By calculating the factors, we can finally derive all
level attributes are summarized in Table 2. users’ stress states over different weeks.
In the following, we will describe the details of the CNN
5 M ODEL F RAMEWORK with CAE and PFG used in the architecture that tackles
the tweet series with cropped modalities and leverages the
Two challenges exist in psychological stress detection. 1) social interaction information between users, respectively.
How to extract user-level attributes from user’s tweeting series
and deal with the problem of absence of modality in the tweets? 5.2 Learning Aggregated Attributes From Tweet Series
2) How to fully leverage social interaction, including interaction
To aggregate user-level attributes, we need to face two major
content and structure patterns, for stress detection? To tackle
challenges: (1) Missing modality, e.g., tweets with only text
these challenges, we propose a novel hybrid model by
but no picture AND (2) How to generate a distributed and
combining a factor graph model with a convolutional neural
modality-invariant representation for each tweets.
network (CNN), since CNN is capable of learning unified
To solve above challenges in cross-media tweet data, we
latent features from multiple modalities, and factor graph
use a cross auto-encoder (CAE) [28] to learn the modality-
model is good at modeling the correlations. In this section,
invariant representation of each single tweet with different
we will first introduce the architecture of our model, and
modalities. Denoting the text, visual, and social attributes of
then describe the details of each part of the proposed model.
a tweet by vT , vI , and vS , the CAE is formulated as follows:

5.1 Architecture u = f (wT vT + wI vI + wS vS + b)
(1)
Figure 3 shows the architecture of our model. There are three +
(vT , vI , vS ) = f (wu b)
types of information that we can use as the initial inputs, i.e., where u is the modality-invariant representation. wT , wI ,
tweet-level attributes, user-level posting behavior attributes, wS , and b are parameters in the encoder, whereas w T ,
and user-level social interaction attributes, whose detailed wI , wS , and b are parameters in the decoder. f (·) is the
computation will be described later. We address the solution activation function. We use a sigmoid activation function
through the following two key components: f (z) = 1+exp(−z) 1
in our model. vT , vI , vS are the recon-
• First, we design a CNN with cross autoencoders structed input modalities.
(CAE) to generate user-level interaction content at- The basic idea of CAE is to force the model to reconstruct
tributes from tweet-level attributes. The CNN has missing modalities in the training stage and to learn cross
been found to be effective in learning stationary modalities correlation from the data (e.g. negative words
local attributes for series like images [3], [6] and in text correlate with cool color in pictures). [18] While
audios [30], [48]. training the cross auto-encoder, we use training data that
• Then, we design a partially-labeled factor graph contains all the three modalities. We manually disable the
(PFG) to incorporate all three aspects of user-level visual modalities and/or social interaction 8 modality of
attributes for user stress detection. Factor graph the training data, and require it to reconstruct all three
model has been widely used in social network mod- modalities. We train the CAE with a cropped set of data
eling [10], [39], [44]. It is effective in leveraging social v T , v I , v S that inputs from one or two modalities are absent,
correlations for different prediction tasks. while requiring it to reconstruct all the three.
We use the stochastic gradient descent to train the CAE.
Take the user labeled with a red star in Figure 3 as an Denoting all the parameters in the CAE as θ, the energy
example. We extract attributes from each tweet of the user function is defined as follows.
to form tweet-level attributes as shown in the cylinders. ⎛ ⎞
Different colors represent different modalities and blank 1⎝ 2
J (v T , v T , v S ; θ) = vM − vM ⎠
(white color) represents modalities that are not available in 2 M ∈T,I,S
the tweet. The tweet-level attributes in the cylinder are fed ⎛ ⎞ (2)
to cross autoencoders (CAEs) [28]. The CAEs are embedded λ⎝ 2 2⎠
in a CNN [26], [29] that will integrate attributes from CAEs + wM + w M .
2 m∈T,I,S
into the aggregated user-level content attributes by pooling
each attribute map. The user-level content attributes, user- 8
Different from the social interaction attributes in this paper, the
level posting behavior attributes, and user-level social in- social interaction here is the attribute of a single tweet defined in [28].
teraction attributes together form the user-level attributes. It is simply the mean and variance of interaction numbers of a tweet.

TABLE 2
Summary of user-level attributes. The column “#“ indicates the feature vector length for each type of feature

Category Short Name # Description

The numbers of @-mentions, @-retweets, and @-replies
Social Engagement 3 in weekly tweet postings, indicating one’s social
interaction activeness with friends.
Posting Behavior
The numbers of tweets posted in hours with a 24-
Tweeting time 24
dimensional vector.
Categorize users’ tweets into mainly four types based
on general categories of social media platforms:
(1) Image tweets (tweets containing images);
(2) Original tweets (tweets that are originally authored
and posted by the user);
(3) Information query tweets (tweets that ask questions
Tweeting type 4
or ask for help );
(4) Information sharing tweets (tweets that contain
outside hyperlinks).
We use a 4-dimensional vector of the numbers of tweets
in the above 4 types respectively to quantify the tweeting
type attribute.
Adopt 10 categories from LIWC that are related to
daily life, social events, e.g., personal
pronouns, home, work, money, religion, death,
Tweeting linguistic style 10 health, ingestion, friends, and family. We
extract words from users’ weekly tweet
postings, and use a 10-dimensional vector
of numbers of words in the 10 categories
A 10-dimensional integer vector,
with each value representing the number of words from
Words 10
Content Style social interaction content of users weekly tweet
postings in each word category from LIWC;
A 2-dimensional integer vector with each value
Social Interaction Emoticons 2 representing the number of positive and negative
emoticons (e.g., and ) in tweets.
Stressed Neighbor Count 1 The number of the user’s stressed neighbors.
Strong-tie Count 1 The number of stressed neighbors with strong tie.
Social Inﬂuence Weak-tie Count 1 The number of stressed neighbors with weak tie.
Follower Count 1 The number of the user’s followers.
Fans Count 1 The number of the user’s fans.
Representing the structure distribution of the user’s
interacted friends, where each element refers to the
Social Structure 8
existence of the corresponding structure
in Fig. 6.

y1t+1

h yit , yit+1

xt1 y1t ?

f xt1 , y1t

g y1t , y2t
g y1t , y3t

xt2 y2t

f xt2 , y2t
xt3
y3t

f xt3 , y3t

h yit , yit−1

y1t−1

Fig. 3. Architecture of our model. The model consists of two parts. The ﬁrst part is a CNN. The second part is a FGM. The CNN will generate
user-level content attributes by convolution with CAE ﬁlters as input to the FGM. Take the user labeled with a red star as example. Tweet-level
attributes of the user are processed through a convolution with CAE to form the user-level content attributes. The user-level attributes are denoted
by xti in the left box. Every xti contains three aspects: user-level content attributes, user-level posting behavior attributes, and user-level social
interaction attributes. Data of other users follows the same route. In the FGM, attribute factors connect user-level attributes to corresponding stress
states. Social factors connect the stress state of different users. Dynamic factors connect stress state of a user over time. The output of the user’s
user-level stress state at time t is y1t as highlighted in red, which actually denotes the stress state of the user in weekly period in this paper.

The first term measures the reconstruction accuracy. The Input: a series of time-varying attribute augmented
second term is the weight decay regularization term that network G with stress states on some of the user
prevents parameters in the model from diverging arbitrarily. nodes, learning rate η ;
λ is the regularization weight. Using data with different Output: parameter value θ = {α, {βc }, γ} and full stress
state vector Y ;
modalities as input, the CAE can be trained and learn a
modality-invariant representation u. Randomly initialize Y ;
The attributes of tweets, which come from a user’s Initialize model parameters θ;
repeat
weekly tweets in timeline, form a time series. To model a
Compute gradient ∇α, ∇βc , ∇γ ;
user as a subject of series of tweets, we apply CNN [26] Update α ← α + η × ∇α;
which has large learning capacity, but has much fewer Update βc ← βc + η × ∇βc ;
connections and parameters to learn than similar-size stan- Update γ ← γ + η × ∇γ ;
dard network layers. It focuses on learning stationary local until convergence;
attributes from series like images (pixel series), audio, and Algorithm 1: Learning and inference by factor graph
other time series. We can learn user-level content attributes
from a series of individual tweets in a time series to describe
a user’s stress state over a week. All attributes of tweets in
a time series form a one-Dimensional series. We use an 1- The joint probability has three types of factor functions,
Dimension CNN in our model. corresponding to the intuitions we have discussed.
CAE units are listed in the attribute maps of the CNN. Attribute factor. We use this factor f (xti , yit ) to represent
They connect to a patch of instance. CAE units take patches the correlation between user vi ’s stress state at time t and
with missing modalities and generate modality-invariant her/his attributes xti . More specifically, we instantiate the
attribute maps. The CAE units are used as filters in the 1-D factor by an exponential-linear function:
CNN and convolute over the sequence of tweets to form one 1
feature map. Thus the latent user-level content attributes can f (xti , yit ) = exp αT xti (4)
Zα
be generated from the tweet-level attributes of single tweets. where α is a parameter of the proposed model, and Zα is
Pooling is another important step to summarize attribute a normalization term.
maps into fewer attribute instances. Though different users
have different number of tweets in different weeks, the Dynamic factor. We use this factor f (yit , yit+1 ) to represent
period of time over which the tweets are sampled are the the time correlation between user vi ’s stress state at time t
and t + 1. More specifically, we instantiate the factor by an
same. We simply pool each attribute map into one pooled exponential-linear function:
attribute. There are two commonly used pooling operations:
1
max-pooling and mean-pooling. When max pooling is used, h(yit , yit+1 ) = exp γ T h (yit , yit+1 ) (5)
Zγ
the pooled attribute unit is assigned with the maximal
activation among all units in the attribute map. When mean- where γ is the model parameters for this type of factor,
pooling is applied, the mean of activations of all units in the h (·) is defined as a vector of indicator functions, and Zγ is
attribute map is assigned to the pooled attribute unit. Since the normalization term.
we pool over the period of time rather than a certain number Social factor. We use social factor g(ye ) (where e =
of tweets, we use mean-over-time (MOT) in this paper, (vit , vjt , c) ∈ E t ) to represent the correlation between user
which can be calculated by summing up the activations, vi and vj ’s stress states according to c at time t:
since the tweet instances are sampled in the same length 1
of time intervals. g(ye ) = exp βc T g (yit , yjt ) (6)
Z βc

5.3 Learning Latent Correlations Between Tweet’s Con- where βc is the model parameters for this type of factor,
g (·) is defined as a vector of indicator functions, and Zβc is
tent And Social Interactions
the normalization term.
As the social correlation between users and time-dependent Finally, by combining Eq.4, 5, 6 into Eq.3, the objective
correlation are hard to be modeled using classic classi- function as the log-likelihood of the proposed model is:
fiers such as SVM, we use a partially-labeled factor graph
T
N
T
N
model (PFG), which was first proposed in [39], to incorpo- O= αT xti + γ T h (yit , yit+1 )
rate social interactions and tweets’ content for learning and t=1 i=1 t=1 i=1
(7)
detecting user-level stress states.
T

We define an objective function by maximizing the con- + βcT g (yit , yjt ) − log Z
t=1 e∈E t
ditional probability of users’ stress states Y given a series of
attribute-augmented networks where Z = Zα c∈C Zβc Zγ is the global normalization
term.
G = {Gt = {(V t , E t , X t , Y t )}}, t ∈ {1, . . . , T }
Learning. Learning the predictive model is to estimate a pa-
and V = V 1 = · · · = V T , |V | = N , i.e., Pθ (Y |G). The
factor graph [25] provides a way to factorize the “global” rameters configuration θ = (α, {βc }, γ) from the partially-
probability as a product of “local” factor functions, which labeled dataset and to maximize the log-likelihood objective
makes the maximization simple, i.e., function Eq. 7, i.e., θ∗ = arg maxθ O(θ).

T
N For optimization, we adopt a gradient decent method.
P (Y |G) = f (xti , yit )h(yit , yit+1 ) g(ye ) (3) Specifically, we derive the gradients with respect to each
t=1 i=1 e∈E t
parameter in our objective function of Eq. 7.

1041-4347 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2017.2686382, IEEE
Transactions on Knowledge and Data Engineering
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 8
T N T N
∂O t t
t t
=E f (xi , yi ) − EPα (Y |G) f (xi , yi ) TABLE 3
∂α t=1 i=1 t=1 i=1
Overview of the Weibo-Stress dataset
⎡ ⎤ ⎡ ⎤
∂O T T
non-stressed stressed total
= E⎣ g(ye )⎦ − EPβ (Y |G) ⎣ g(ye )⎦
∂β #tweets 253, 638 239, 038 492, 676
t=1 e∈E t t=1 e∈E t
T N T N #users 12, 230 11, 074 23, 304
∂O #weeks 17, 861 19, 136 36, 997
=E h(yit , yit+1 ) − EPγ (Y |G) h(yit , yit+1 ) #tweets/week* 14.2 12.5 13.3
∂γ t=1 i=1 t=1 i=1 #weeks/user* 1.46 1.73 1.59
(8) #interacting users/week* 5.79 6.99 6.35
N * means average number.
T t t
where in the first equation, E t=1 i=1 f (x ,
i iy ) is
the expectation of the summation of the attribute factor (retweet, comment, favorite), video and etc. Despite these
functions given the data distribution
over Y and G in the user generated contents, user relationship, which takes the
T N t t
training set, and EPα (Y |G) t=1 i=1 f (x ,
i iy ) is the ex- form of following on Sina Weibo, also contains abundant
pectation of the summation of the attribute factor functions information for data analysis. Utilizing above information
given by the estimated model. The other expectation terms and features extracted from multiple modalities, we are able
have similar meanings in the other equation. to investigate users emotions, stresses and opinions.
As the network structure in the real world may contain We then tried to identify the weekly stressed state of
cycles, it is intractable to estimate the marginal probability users. Facing the vast scale of social images, manually
in the second terms of 8. In this work, we adopt Loopy Belief labeling is powerless. Instead, we use tags and comments
Propagation (LBP) [33] to calculate the marginal probability for automatic image labeling, which is a common method
of P (Y ) and compute the expectation terms. The learning in previous work. [20], [21], [46] This is done by searching
process can then be described as an iterative algorithm. Each
iteration contains two steps. Firstly, we call LBP to calculate for tweets containing patterns like “I feel stressed this week”
marginal distributions of unknown variables Pα (Y |G). Sec- and “I feel stressed so much this week”, which are used to
ondly, we update α, β , γ with the learning rate η by Eq.9 The indicate that the users are stressed. The weeks containing
learning algorithm terminates when it reaches convergence. such sentence patterns are labeled as “stressed” weeks.
∂O Similarly, we identify “non-stressed” weeks of users by
θnew = θold + η (9) searching for tweets with patterns like “I feel relaxed” and “I
∂θ
feel non-stressed”. These sentence patterns have been shown
Detection. With the estimated parameter θ, we can now
to have high precision against user-assigned psychological
assign the value of unknown labels Y by looking for a label state labels validated by online surveys in weibo [29].
configuration that will maximize the objective function, i.e., In this way, we collected over 19,000 weeks of tweets
Y ∗ = arg max O(Y |G, θ) (10) that are labeled as stressed, and over 17,000 weeks of non-
stressed users’ tweets. There are 492,676 tweets from 23,304
In this paper, we use a max-sum algorithm [31] to solve users in total. We use this dataset for experiments, analysis
this problem. and further in-depth studies, which is represented by DB1
in this paper. Details of the dataset are shown in Table 3.
6 E XPERIMENTS Dataset DB2: We verified the reliability of the above
In this section, we will present the effectiveness and effi- ground truth labeling method through dataset DB2 in Ta-
ciency of our hybrid model on user-level stress detection. ble 4. It is a small dataset collected from the users who
have shared the score of a psychological stress scale PSTR10
designed by psychologists via Weibo. Guided by the rules
6.1 Dataset Collection of the PSTR scale, a user is taken as stressed when the score
To conduct observations and evaluate our succecive model, is larger than 80, otherwise non-stressed. We thus crawled
we first collect a set of datasets using different labeling the scores posted by users, and used the scores as ground
methods, which are listed as following: truth label for the set of tweets in +-3-day window.
Dataset DB1: It is a challenge to construct a dataset with Dataset DB3 and DB4: To further test our method, we
reliable ground truth labels from large-scale noisy social collected two more datasets from Tencent Weibo (DB3) and
media data. The data crawled from social platforms is usu- Twitter (DB4). They are again labeled using the sentence
ally massive, thus manual labeling methods are not feasible pattern labeling method as described above for DB1. In
due to the uncontrollable cost and quality. To solve this particular, as social platforms of different languages, Weibo
problem, we employed a sentence pattern labeling method and Twitter have many differences. [51]. For example, their
to automatically extract labeled data from the crawled large- top topics differs very much. Thus, experiments on Twitter
scale social media data. We first crawled 350 million tweets can validate the universality of our method. The details of
data via Sina Weibo’s REST APIs9 from Oct. 2009 to Oct. the two datasets are presented in Table 4.
2012. Sina weibo, as the biggest microblog website in China,
provides users an open online platform for information
6.2 Experimental Setup
sharing, communication and obtaining. Similar to Twitter
and Facebook, users on Sina Weibo can post contents with In the following experiments, we first train and test our
multiple modalities, including text, image, social action model on the large-scale Sina Weibo dataset DB1. We then
9 10
http://open.weibo.com http://types.yuzeli.com/survey/pstr50

TABLE 4
Details of other Datasets

Platform Stress label Number of tweets Number of users Number of weeks Tweets per week
stressed 1, 459 98 98 14.9
DB2:Sina Weibo
non-stressed 1, 845 112 112 16.5
(2010.2-2011.9)
summary 3, 304 210 210 15.7
stressed 138, 570 7, 845 8, 974 15.4
DB3:Tencent Weibo
non-stressed 172, 585 8, 239 9, 976 17.3
(2011.11-2013.3)
summary 311, 155 16, 084 18, 950 16.4
stressed 54, 748 4, 905 6, 081 9.0
DB4:Twitter
non-stressed 75, 357 4, 018 6, 545 11.5
(2009.6-2009.12)
summary 130, 105 8, 923 12, 626 10.3

test our model on the other 3 datasets to show effectiveness TABLE 5

of the proposed model on different data sources or different Comparison of efficiency and effectiveness using different models (%).
ground truth labeling methods. For all of our analysis,
we use 5-fold cross validation, with over 10 randomized Method Acc. Rec. Prec. F1 CPU time
LRC 76.18 87.94 78.58 83.00 39.43s
experimental runs. SVM 72.58 87.39 75.16 80.82 ≈10min
RF 77.73 89.63 79.35 84.18 67.71s
Comparison Methods. We compare the following classifi- GBDT 79.75 82.99 85.90 84.43 262.86s
cation methods for user-level psychological stress detection FGM 91.55 96.56 90.44 93.40 ≈20min
with our FGM+CNN model (denoted as FGM here).
different models based on the Weibo-Stress dataset. In
• Logistic Regression (LRC) [19]: it trains a logistic re- this experiment, we used all the three attributes described
gression classification model and then predicts users’ in previous section: user-level social interaction attributes,
labels in the test set. user-level posting behavior attributes and user-level con-
• Support Vector Machine (SVM) [5]: it is a popular tent attributes generated from the tweet-level attributes by
and binary classifier that is proved to be effective CNN+CAE. Table 5 shows the experimental results. We see
on a huge category of classification problems. In our that FGM gains superior results against the comparative
problem we use SVM with RBF kernel. methods, which verifies that our proposed model can ef-
• Random Forest (RF) [42]: it is an ensemble learn- fectively leverage the social interaction and social structure
ing method for decision trees by building a set of attributes for stress detection. Compared with the results in
decision trees with random subsets of attributes and [29], which also aims at user-level stress detection based on
bagging them for classification results. social media data sources, our proposed model improves
• Gradient Boosted Decision Tree (GBDT) [13]: it the detection performance by up to 9% on F1-score. These
trains a gradient boosted decision tree model with results demonstrate the feasibility of stress detection via the
features associated with each user. brand new information source of social interactions, and
• Deep Neural Network (DNN) [29] for user-level that our proposed model can significantly enhance the per-
stress detection: it is proposed to deal with the formance by leveraging the social interaction information.
problem of user-level stress detection problem with We further perform t-tests and all the p-values are ≤ 0.01,
a convolutional neural network (CNN) with cross indicating that the improvements of our proposed models
autoencoders. This is the real baseline method that over the comparison methods are statistically significant.
we can compare our proposed model with.
Comparison of Model Efficiency. To evaluate the efficiency
We employ scikit-learn11 for the above methods.
of the aforementioned methods, we compare the CPU time
Evaluation Measures. For a fully investigation of the of training each model. The comparison results are also
proposed methods, we consider the following aspects: shown in Table 5. Overall, all methods have good efficiency
performance, and the running time of different methods
• Effectiveness. We evaluate the detection perfor-
ranges from seconds to minutes. FGM results in a slightly
mance of our model and comparison methods
lower but better performance compared to other methods.
in terms of Accuracy (Acc.), Recall (Rec.), Preci-
sion (Prec.) and F1-Measure (F1) [2]. Factor Contribution Analysis. The definition of factors is
• Efficiency. We evaluate efficiency of the methods by important to the performance of the Factor Graph Model.
comparing the CPU time of training each model. All We have three types of factors in our model, i.e., attribute
experiments are performed on an x64 machine with factor, social factor, and dynamic factor. To analyze the
2.9GHz intel Core i7 CPU and 8GB RAM. impact of different factors in our model, we compare the
detection performance with different combinations of fac-
tors in this experiment, as shown in Figure 4(a). Specifically,
6.3 Experimental Results on DB1
we first use all the three factors, denoted as FGM, then
Comparison of Detection Performance. To evaluate the we remove the following factors respectively: social factor,
effectiveness of our model, we first conduct a test using dynamic factor and both of them, denoted as FGM-S, FGM-
D and FGM-S-D We see that the worst performance is
11
http://scikit-learn.org achieved if we incorporate only the attribute factor. How-

95 100 100 95
FGM Accuracy(%)
FGM-S 95
90 F1-score(%)
90 FGM-D 80 90
FGM-S-D
F1-score(%)

F1-score(%)

F1-score(%)
90 70
85 85
85 60

80 50 80
80
40
75 750% 300 75
10% 20% 30% 40% 50% 60% 70% 80% 500 1000 1500 2000 2500 3000 3500 4000 DNN-1 DNN-2 DNN-3 DNN-4
(a) Factor Contribution (b) Data Scale Analysis (c) Convergence Analysis (d) DNN Layer Analysis
Fig. 4. Experiment results analysis from various perspectives. (a) Attribute contribution analysis; (b) Factor contribution analysis; (c) Results of
detection performance with different training data scales; (d) Convergence Analysis of FGM.

100

95
SVM FGM CNN-FGM in setting up DNN model. Shallow networks result in trivial
90
model that cannot catch any underlying correlation in data,
85 whereas too deep networks lead to over-complex model
F1-score(%)

80
which is difﬁcult to tune and may suffer from problems
75
like over-ﬁtting. To choose an appropriate DNN model
70

65
for classification, we test DNN with different number of
60 layers. Figure 4(d) summarizes the experiment results. It is
55
B UIC UIS l clear that 2-layer is not sufficient for the model to achieve
T
UP PB UIC UIS UIC UIS UIS UIS UIC UIS UIS Al
T+U T+ T+ UPB+ UPB+ UIC+ UIC+ UPB+ UPB+ UIC+ a satisfactory result. 3-layer model improve significantly
B+ + T+ T+
UP T while 4-layer model reaches the peak. 5-layer model does
not get better result. This is mainly because at 5-layer the
Fig. 5. Experiment results analysis of different attribute combinations network may be too large that it cannot be tuned well with
on different models, with T, UPB, UIC, and UIS representing Tweet-
level attributes, User-level Posting Behavior attributes, User-level social the available data and within a feasible training time.
Interaction Content attributes and User-level social Interaction Structure
attributes respectively. For example, ‘UIC+UIS’ here means a combi- Attribute Contribution Analysis. As described in Section 4,
nation of User-level social Interaction Content attributes and User-level we have defined several set of tweet-level and user-level
social Interaction Structure attributes. attributes from a single tweet’s content as well as users’
ever, integrating attribute factor with social or dynamic posting behaviors and social interactions in a weekly pe-
factor both improve the performance, revealing that both of riod. To evaluate the contribution of different attributes
the two factors are effective for stress detection. Specifically, and compare the effectiveness of our model of leveraging
incorporating social factor significantly improves the detec- different attributes, we compared the proposed model with
tion performance to around 91% on accuracy, indicating that other existing models by using different combinations of
the social factor is extremely effective. The best detection attributes as input. As described in Section 4, the proposed
performance is observed when using all three types of attributes are categorized into four groups: tweet-level at-
factors. tributes, user-level posting behavior attributes, user-level
social interaction content attributes and user-level social
Training Data Scale Analysis. To evaluate the data scala- interaction structure attributes, denoted as T, UPB, UIC, and
bility of the proposed model, we try to train the model with UIS respectively. We compare the detection performance of
different scale of training data, and compare the final detec- the proposed CNN+FGM model with SVM and CNN with
tion performance in F1-score. In this test, we use all the three traditional autoencoder, with all the possible combinations
attributes as input. Figure 4(b) shows the trend of detection of these four set of attributes. For the SVM with the tweet-
performance with different proportions of training data. It level attributes, we simply take the average of the feature
is clear that when using only 1% of all training data, our vectors from a user’s weekly tweets.
model fails to achieve meaningful detection performance. The results of this experiment are shown in Figure 5. We
When adopting approximately 30% of all training data, our see that all the models achieve the best detection perfor-
model can obtain an equally competitive performance of mance when utilizing all the three set of attributes. When
around 93% compared with that when using 50% of training using only the tweet-level features, the detection perfor-
data. Moreover, the performance keeps increasing given mance of the proposed model and the DNN model drops
more training data. These results verify the scalability of to around 86% and 82% respectively in F1-score, which is
our model on large-scale real-world social media datasets. acceptable. While for SVM, the detection performance drops
to around only 70%, which is poor for a binary classification.
Convergence Analysis. We further investigate the conver-
This result demonstrates the effectiveness of the feature
gence of the learning algorithm for FGM, and Figure 4(c)
aggregation of CNN, which is much better than simply
presents the F1-score with increasing number of iterations.
summarizing the feature vectors manually.
We see that the algorithm converges within around 2000
iterations, which is rapid enough for us to conduct efficient Figure 5 also shows the effectiveness of different at-
model training on large scale datasets in practice. tributes. We can see that by using only user-level attributes,
the detection performance of all the models drops dras-
Impact of size of network. Size of network is a critical issue tically compared to that using only tweet-level attributes,

TABLE 6 1.0
Comparison of results using different modalities. streeed
non-stressed
Text Text + visual Text + Social All 0.8
Accuracy 0.8713 0.8761 0.8628 0.9155
F1-score 0.8794 0.8865 0.8711 0.9340 0.6

Proportion
which shows the importance of the tweet-level attributes.
0.4
By combing different types of user-level attributes, the
detection performance improves by around 3-8% in F1- 0.2
score, showing that the user-level attributes are supple-
mentary to each other. Meanwhile, by combining the user- 0.0
liy nd mo ect ure sad death anx anger gemo
fami frie pose aff leis
level attributes with tweet-level attributes, the detection
ne
performance improves up to 10-20% in F1-score. This result
Fig. 6. Distribution of stress states (stressed and non-stressed) over
indicates that the user-level attributes are great supplements different word categories from tweets’ comments and retweets. Here we
to tweet-level attributes. show 10 most widely used word categories in our data set.
When using only two set of attributes, the detection per-
formance drops to around 91% in F1-score. In case of using DB4 from Twitter. We also test against the Twitter
sole attributes, we see that using solely user-level social dataset. We still use the attribute extractor trained with large
interaction attributes gets the best detection performance of scale Sina Weibo dataset and only finetune the network with
around 90% in F1-score, as compared to the other attributes. Twitter dataset in 5-fold. The accuracy is 77.43% and F1-
This reveals that the proposed user-level social interaction score is 0.8224. One reason for this modest result is that users
attributes are quite effective for stress detection. in Twitter dataset and Sina Weibo dataset come from differ-
ent language and culture background, so that the language
Impact of different modalities in content attributes: patterns and sentimental signals from these two different
Tweets content come with multiple modalities. To eval- language environments can be different, thus the attribute
uate the contribution of each data modality, we conduct extractor trained with large scale Sina Weibo dataset may
experiments with different combination of attributes. Since not be fully functional for Twitter datasets. Nevertheless,
text is the necessary part of a tweet, we test using solely we still achieved acceptable performance in Twitter dataset,
text attributes, and the two combinations of text and vi- which implies that the basic stress patterns between social
sual attributes, and text and social attributes, as well as relations can be transferred in between different language
using all attributes. The results are shown in Table 6. It environments. Another factor could be that the scale of this
is interesting to note that using only text attribute could dataset is rather small. Subjects in the Twitter dataset are on
achieve rather high performance. Simply combining visual the order of 10% than that in large-scale Sina Weibo dataset.
or social attributes with text attributes may even reduce the We look into the collected data and find that, by coincidence,
performance, especially the social attributes. This trend is all tweets in this dataset have no social activity. We conjec-
even more obvious when both types of attributes (content ture this is also one of the causes of the unsatisfactory result.
and posting behavior) are used. Nevertheless, using all
attributes together outperforms that using only the text
attributes; and the highest performance is observed when 7 S TUDIES OF S OCIAL I NTERACTION
using all attribute and working with both types of attributes.
We have presented the experimental results on stress detec-
6.4 Results on Other Datasets tion in the previous section, while in the setting of social
We further evaluate our model on other datasets, DB2-DB4, networks, it would be helpful to further analyze how a
as shown in Table 4, to show that our model is universally user’s stress status is developed and how they affect each
applicable. For these experiments, we use all the proposed other. To do so, we try to conduct several studies on DB1 to
attributes with MOT pooling, and a 4-layer DNN model. offer insights on how social interactions contribute to user
stress and the task of stress detection from the following
DB2 from Sina Weibo with PSTR label.
aspects:
We use a matured model trained with large scale Sina
(1) Content. How are users’ social interaction contents (e.g.,
Weibo dataset, and then test it against another set of subject
language used) related to users’ stress states?
independently sampled from Sina Weibo. For the test set,
we collect weekly tweets from the users that have shared (2) Structure. Compared to non-stressed users, do stressed
the score of a psychological stress scale with 50 items via users show different structural diversity patterns when they
Sina Weibo. Detection result shows that the test accuracy is behave in social networks? Do differences of social influence and
84.26% and F1-score is 0.8785, which demonstrates that the strong/weak ties exist between stressed and non-stressed users?
overall model is consistent and the sentence pattern based
ground truth labeling method is reliable. 7.1 Content
DB3 from Tencent Weibo. We test on data collected from Content of social interaction refers to the content of tweets’
another major Chinese social media platform. For this test, comments and retweets, including text, emoticons, and
we use the attribute extractor trained with large scale Sina punctuation marks. Based on a widely used psychologi-
Weibo dataset and only finetune the network with Twitter cal dictionary LIWC2007 [40], we extract emotional words
dataset in 5-fold. The accuracy is 86.18% and F1-score is from the interaction content of tweets, and categorize
0.8832 which demonstrate the capability of the model. the extracted words into corresponding groups defined in

LIWC2007. We compare the frequencies of different word

categories between stressed and non-stressed users.
Figure 6 shows the comparison results of the most

widely used word categories in our data set, we observe

that there is an obvious difference in interaction contents
between stressed and non-stressed users. That is, interac-

tion contents of stressed users’ tweets contains much more

words from categories like death, sadness, anxiety, anger,
and negative emotion, while non-stressed users’ tweets
contain more words from categories like friends, family,
affection, leisure, and positive emotion.

7.2 Structure
Fig. 7. Distribution of stress states (stressed and non-stressed) over
To examine structure properties (i.e., social influence and different social structures. The dot represents a friend of the user, and
strong/weak tie) of (non)stressed users, we use risk ratio the line represents the connection of friends.
(RR) to measure the correlation between users’ stress states
7.2.2 Social Influence
and different structural attributes. Risk ratio is an effective
measurement widely used in the statistical analysis and Social influence is an important factor that governs the
relevant fields. The risk ratio of a stressed state, associated dynamics of social networks. The principle of social influ-
with a structural attribute a, is calculated as follows: ence [22] suggests that users tend to change their behaviors
P (stressed user has attribute a) to match their friends’ behaviors. In this study, we try
RR(a) = . to examine whether users’ stress states will be influenced
P (stressed user does not have attribute a)
(11) by their neighbors’ states by looking at the probability of
A larger risk ratio implies that users with attribute a are a user’s stress state when he/she has different types of
more likely to be stressed. In this section, we investigate relationships with other stressed users. As for the stress state
representative sociology theories, and quantitatively ana- labeling, all users including friends are labeled using the
lyze the correlations between users’ stress states and fun- sentence pattern method described in previous section.
damental social concepts, so as to examine how and why a Figure 8(a) shows the probability that a user being
user’s stress state is developed and affected by other users. stressed, conditioned on the number of stressed neighbors
that the user has in the social network. We can see that being
7.2.1 Structural Diversity stressed is a mutually correlated behavior. In particular,
We are interested in whether stressed and non-stressed the chance that a none-stressed user becoming stressed
users have any structural difference in respective friends’ increases to three times higher for those with stressed neigh-
connection. In sociology, social structure refers to a society’s bors than for those without. Another trend observed from
framework, consisting of various relationships among peo- Figure 8(a) is that the likelihood of a user becoming stressed
ple, as well as groups that direct and set limits on human increases with the number of stressed neighbors.
behaviors. In social networks, direct connections (following
0.09

or followed) of users that interact with each other via com- 0.08
0.20
Strong Tie-5

Strong Tie-3

ments and retweets also form a kind of social structure. For 0.07
0.15
Weak Tie
Probability

Probability

this in-depth study, we select top four users with the most 0.06

x9
x4
frequent interactions from users’ weekly tweet postings, 0.05 0.10

where four is adopted because this is the minimum number

0.04

x2
0.05
0.03

of nodes required to produce structural combinations (10 0.02

combinations), so as to calculate the probability of each 1 2 3 4

#Stressed Neighbors
5
0.00
1 2 3 4

#Stressed Neighbors
5

combination, and incorporating more nodes would make (a) (b)

the calculation combinatorial expensive. We measure the
Fig. 8. Social influence and Social tie analysis. (a) Variation trend of
connection of the interacting users by the following link, probability of a user being stressed when she/he has different number
that is, if A is following or followed by B, then A and B of stressed neighbors; (b) Variation trend of probability of a user being
are connected, and cliques made up of different nodes are stressed when she/he has different number of stressed neighbors with
strong/weak ties.
treated the same. We compare the proportion of different so-
cial structures of interacting users to measure the structural 7.2.3 Strong/Weak Tie
diversity. The results in Figure 7 clearly show us that struc- Strong/Weak Tie [17] is one of the most basic principles in
tural differences do exist between stressed and non-stressed social network theories. We classify the constructed social
users. The number of social structures of sparse connection relationships into strong or weak ties by the number of
(i.e. with no delta connections) of stressed users is around times that two users interact with each other via comment,
14% higher than that of non-stressed users, indicating that @-mention, retweet, or like in a week. In our work, we tried
the social structure of stressed users’ friends tends to be less different values for the threshold and finally chose three by
connected and less complicated, compared to that of non- cross-validation. If two users interact with each other more
stressed users. This phenomenon has also been reported by than three times, we call the relationship a strong tie, and
the current psychological research result that stressed users otherwise a weak tie. This definition of user ties is adopted
are more likely to be socially less active [7]. as the standard treatment in the research of social network

analysis [17], so as to capture the most recent user rela- [8] Glen Coppersmith, Craig Harman, and Mark Dredze. Measuring
tionships in a shifting environment. Figure 8(b) illustrates post traumatic stress disorder in twitter. In Proceedings of the
International Conference on Weblogs and Social Media, pages 579–582,
the results. We can see that strong ties indeed have strong 2014.
influence on users’ stress states, and the influence of weak [9] Rui Fan, Jichang Zhao, Yan Chen, and Ke Xu. Anger is more
ties is relatively weak. For example, when a user has three influential than joy: Sentiment correlation in weibo. PLoS ONE,
stressed strong-tie connections, the probability that the user 2014.
[10] Zhanpeng Fang, Xinyu Zhou, Jie Tang, Wei Shao, A.C.M. Fong,
will become stressed increases to 13%, more than twice as Longjun Sun, Ying Ding, Ling Zhou, , and Jarder Luo. Modeling
high as for a user with three stressed weak-tie connections. paying behavior in game social networks. In In Proceedings of the
Twenty-Third Conference on Information and Knowledge Management
Summary. Based on the experimental results and analyses (CIKM’14), pages 411–420, 2014.
we know that: 1) users’ stress states are not only revealed [11] Golnoosh Farnadi, Geetha Sitaraman, Shanu Sushmita, Fabio
in their own tweets, but also affected by the contents of Celli, Michal Kosinski, David Stillwell, Sergio Davalos,
their social interactions, including commenting on and re- Marie Francine Moens, and Martine De Cock. Computational
personality recognition in social media. User Modeling and User-
tweeting others’ tweets; and 2) users’ stress states are re- Adapted Interaction, pages 1–34, 2016.
vealed by the structure of their social interactions, including [12] Eileen Fischer and A. Rebecca Reuber. Social interaction via new
structural diversity, social influence, and strong/weak ties. social media: (how) can interactions on twitter affect effectual
thinking and behavior? Journal of Business Venturing, 26(1):1–18,
These insights quantitatively prove the necessity and effec- 2011.
tiveness of combining social interactions for stress detection. [13] Jerome H. Friedman. Greedy function approximation: A gradient
boosting machine. Annals of Statistics, 29(5):1189–1232, 1999.
[14] Rui Gao, Bibo Hao, He Li, Yusong Gao, and Tingshao Zhu.
8 C ONCLUSION Developing simplified chinese psychological linguistic analysis
dictionary for microblog. pages 359–368, 2013.
In this paper, we presented a framework for detecting users’ [15] Johannes Gettinger and Sabine T. Koeszegi. More Than Words: The
psychological stress states from users’ weekly social media Effect of Emoticons in Electronic Negotiations.
data, leveraging tweets’ content as well as users’ social inter- [16] Jennifer Golbeck, Cristina Robles, Michon Edmondson, and Karen
Turner. Predicting personality from twitter. In Passat/socialcom
actions. Employing real-world social media data as the ba- 2011, Privacy, Security, Risk and Trust, pages 149–156, 2011.
sis, we studied the correlation between user’ psychological [17] Mark S. Granovetter. The strength of weak ties. American Journal
stress states and their social interaction behaviors. To fully of Sociology, 1973.
leverage both content and social interaction information of [18] Quan Guo, Jia Jia, Guangyao Shen, Lei Zhang, Lianhong Cai, and
Zhang Yi. Learning robust uniform features for cross-media social
users’ tweets, we proposed a hybrid model which combines data by using cross autoencoders. Knowledge Based System, 102:64–
the factor graph model (FGM) with a convolutional neural 75, 2016.
network (CNN). [19] David W. Hosmer, Stanley Lemeshow, and Rodney X. Sturdivant.
In this work, we also discovered several intriguing phe- Applied logistic regression. Wiley series in probability and mathemat-
ical statistics, 2013.
nomena of stress. We found that the number of social struc- [20] Sung Ju Hwang. Discriminative object categorization with exter-
tures of sparse connection (i.e. with no delta connections) nal semantic knowledge. 2013.
of stressed users is around 14% higher than that of non- [21] Sepandar D. Kamvar. We feel fine and searching the emotional
web. In In Proceedings of WSDM, pages 117–126, 2011.
stressed users, indicating that the social structure of stressed
[22] Herbert C. Kelman. Compliance, identification, and internal-
users’ friends tend to be less connected and less complicated ization: Three processes of attitude change. general information,
than that of non-stressed users. These phenomena could be 1(1):51–60, 1958.
useful references for future related studies. [23] Shigenobu Kobayashi. The aim and method of the color image
scale. Color research & application, 6(2):93–107, 1981.
[24] Novak P Kralj, J Smailovi, B Sluban, and I Mozeti. Sentiment of
emojis. Plos One, 10(12), 2015.
R EFERENCES [25] Frank R Kschischang, Brendan J Frey, and H-A Loeliger. Factor
[1] Andrey Bogomolov, Bruno Lepri, Michela Ferron, Fabio Pianesi, graphs and the sum-product algorithm. Information Theory, IEEE
and Alex Pentland. Daily stress recognition from mobile phone Transactions on, 47(2):498–519, 2001.
data, weather conditions and individual traits. In ACM Interna- [26] Yann LeCun and Yoshua Bengio. Convolutional networks for
tional Conference on Multimedia, pages 477–486, 2014. images, speech, and time series. The handbook of brain theory and
[2] Chris Buckley and Ellen M Voorhees. Retrieval evaluation with in- neural networks, 3361, 1995.
complete information. In Proceedings of the 27th annual international [27] Kathy Lee, Ankit Agrawal, and Alok Choudhary. Real-time
ACM SIGIR conference on Research and development in information disease surveillance using twitter data: demonstration on flu and
retrieval, pages 25–32, 2004. cancer. In Proceedings of ACM SIGKDD international conference on
[3] Xiaojun Chang, Yi Yang, Alexander G Hauptmann, Eric P Xing, Knowledge discovery and data mining, pages 1474–1477, 2013.
and Yao-Liang Yu. Semantic concept discovery for large-scale [28] H. Lin, J. Jia, Q. Guo, Y. Xue, J. Huang, L. Cai, and L. Feng. Psy-
zero-shot event detection. In Proceedings of International Joint chological stress detection from cross-media microblog data using
Conference on Artificial Intelligence, pages 2234–2240, 2015. deep sparse neural network. In proceedings of IEEE International
[4] Wanxiang Che, Zhenghua Li, and Ting Liu. Ltp: A chinese Conference on Multimedia & Expo, 2014.
language technology platform. In Proceedings of International Con- [29] H. Lin, J. Jia, Q. Guo, Y. Xue, Q. Li, J Huang, L. Cai, and L. Feng.
ference on Computational Linguistics, pages 13–16, 2010. User-level psychological stress detection from social media using
[5] Chih chung Chang and Chih-Jen Lin. Libsvm: a library for sup- deep neural network. In Proceedings of ACM Int. Conference on
port vector machines. ACM TRANSACTIONS ON INTELLIGENT Multimedia, 2014.
SYSTEMS AND TECHNOLOGY, 2(3):389–396, 2001. [30] Li Liu and Ling Shao. Learning discriminative representations
[6] Dan C Ciresan, Ueli Meier, Jonathan Masci, Luca Maria Gam- from rgb-d video data. In Proceedings of International Joint Confer-
bardella, and Jürgen Schmidhuber. Flexible, high performance ence on Artificial Intelligence, pages 1493–1500, 2013.
convolutional neural networks for image classification. In Proceed- [31] H-A Loeliger. An introduction to factor graphs. Signal Processing
ings of International Joint Conference on Artificial Intelligence, pages Magazine, IEEE, 21(1):28–41, 2004.
1237–1242, 2011. [32] Jana Machajdik and Allan Hanbury. Affective image classification
[7] Sheldon Cohen and Thomas A. W. Stress, social support, and the using features inspired by psychology and art theory. In Proceed-
buffering hypothesis. Psychological Bulletin, 98(2):310–357, 1985. ings of the international conference on Multimedia, pages 83–92, 2010.

[33] Kevin P Murphy, Yair Weiss, and Michael I Jordan. Loopy belief Huijie Lin is currently a Ph.D. candidate of Tsinghua University.
propagation for approximate inference: An empirical study. In
Proceedings of the Fifteenth conference on Uncertainty in artificial Jia Jia is an associate professor in Department of Computer
intelligence, pages 467–475, 1999. Science and Technology, Tsinghua University. She got bachelor degree
[34] Cristian Danescu niculescu mizil, Lillian Lee, Bo Pang, and Jon at Tsinghua University in 2003, and received her Ph.D. degree from
Kleinberg. Echoes of power: Language effects and power differ- Tsinghua University in 2008. Her main research interest is social
ences in social interaction. eprint arXiv:1112.3670, 2011. affective computing and human computer speech interaction. She is
[35] Liqiang Nie, Yi-Liang Zhao, Mohammad Akbari, Jialie Shen, and serving as the secretary-general of Professional Committee of Speech
Tat-Seng Chua. Bridging the vocabulary gap between health seek- in Chinese Information Processing Society, and also a committee
ers and healthcare knowledge. Knowledge and Data Engineering, member of Multimedia Federation in China Society of Image and
IEEE Transactions on, 27(2):396–409, 2015. Graphic.
[36] Federico Alberto Pozzi, Daniele Maccagnola, Elisabetta Fersini,
and Enza Messina. Enhance user-level sentiment analysis on Jiezhong Qiu is currently an Ph.D student of Tsinghua University.
microblogs with approval relations. In AI* IA 2013: Advances in
Artificial Intelligence, pages 133–144. 2013. Yongfeng Zhang is a Postdoc Research Associate in University
[37] Neumann R and Strack F. ”mood contagion”: the automatic of Massachusetts Amherst, US.
transfer of mood between persons. Journal of Personality and Social
Psychology, pages 211–223, 2000. Lexing Xie received the B.S. degree from Tsinghua University,
Beijing, China, in 2000, and the M.S. and Ph.D. degrees from Columbia
[38] Chenhao Tan, Lillian Lee, Jie Tang, Long Jiang, Ming Zhou,
University, in 2002 and 2005, respectively, all in electrical engineering.
and Ping Li. User-level sentiment analysis incorporating social
She is a Lecturer in the Research School of Computer Science at
networks. In Proceedings of the SIGKDD international conference on
the Australian National University; she was with the IBM T.J. Watson
Knowledge discovery and data mining, pages 1397–1405, 2011.
Research Center, Hawthorne, NY from 2005 to 2010. Her recent
[39] Wenbin Tang, Honglei Zhuang, and Jie Tang. Learning to infer research interests are in multimedia mining, machine learning and
social ties in large networks. In Machine Learning and Knowledge social media analysis. Dr. Xie has won several awards: the best
Discovery in Databases, pages 381–397. 2011. conference paper award in IEEE SOLI 2011, the best student paper
[40] Y. R. Tausczik and J. W. Pennebaker. The psychological meaning awards at JCDL 2007, ICIP 2004, ACM Multimedia 2005 and ACM
of words: Liwc and computerized text analysis methods. Journal Multimedia 2002.
of Language and Social Psychology, 29(1):24–54, 2010.
[41] Mike Thelwall, Kevan Buckley, Georgios Paltoglou, Di Cai, and Jie Tang is an associate professor at the Department of Computer
Arvid Kappas. Sentiment strength detection in short informal text. Science and Technology, Tsinghua University. His main research
Journal of the American Society for Information Science and Technology, interests include data mining algorithms and social network theories.
61(12):2544–2558, 2010. He has been visiting scholar at Cornell University, Chinese University
[42] Svetnik V. Random forest: a classification and regression tool for of Hong Kong, Hong Kong University of Science and Technology, and
compound classification and qsar modeling. Journal of Chemical Leuven University. He has published over 100 research papers in major
Information and Computer Sciences, 43(6):1947–1958, 2003. international journals and conferences including: KDD, IJCAI, AAAI,
[43] Ben Verhoeven, Walter Daelemans, and Barbara Plank. Twisty: A ICML, WWW, SIGIR, SIGMOD, ACL, Machine Learning Journal, TKDD,
multilingual twitter stylometry corpus for gender and personality and TKDE.
profiling. In Proceedings of the Tenth International Conference on
Language Resources and Evaluation LREC, pages 1632–1637, 2016. Ling Feng is a professor of computer science and technology at
[44] Chi Wang, Jie Tang, Jimeng Sun, and Jiawei Han. Dynamic Tsinghua University, Beijing. Her research interests include context-
social influence analysis through time-dependent factor graphs. aware data management toward ambient intelligence, knowledge-based
Advances in Social Networks Analysis and Mining (ASONAM), 2011 information systems, data mining and warehousing, and distributed
International Conference on, pages 239 – 246, 2011. object-oriented database management systems. She has published
[45] Xiaohui Wang, Jia Jia, Peiyun Hu, Sen Wu, Jie Tang, and Lianhong more than 150 scientific articles in high-quality international conferences
Cai. Understanding the emotional impact of images. In Proceedings or journals, and received the 2004 Innovational VIDI Award by the
of the 20th ACM International Conference on Multimedia, MM ’12, Netherlands Organization for Scientific Research, the 2006 Chinese
pages 1369–1370, 2012. ChangJiang Professorship Award by the Ministry of Education, and the
[46] Lexing Xie and Xuming He. Picture tags and world knowledge: 2006 Tsinghua Hundred-Talents Award.
learning tag relations from visual semantic sources. In ACM
Multimedia Conference, pages 967–976, 2013. Tat-Seng Chua joined the National University of Singapore, Singapore,
[47] Yuanyuan Xue and et al. Detecting adolescent psychological in 1983, and spent three years as a Research Staff Member with
pressures from micro-blog. Health Information Science. Springer the Institute of Systems Science, National University of Singapore.
International Publishing, pages 83–94, 2014. He was the Acting and Founding Dean of the School of Computing,
National University of Singapore, from 1998 to 2000. He is currently
[48] Jian Bo Yang, Minh Nhut Nguyen, Phyo Phyo San, Xiao Li Li,
the KITHCT Chair Professor with the School of Computing, National
and Shonali Krishnaswamy. Deep convolutional neural networks
University of Singapore. His research interests include multimedia
on multichannel time series for human activity recognition. In
information retrieval, multimedia question answering, and the analysis
Proceedings of International Joint Conference on Artificial Intelligence,
and structuring of user-generated contents.
pages 3995–4001, 2015.
Dr. Chua has organized and served as a Program Committee Mem-
[49] Yang Yang, Jia Jia, Shumei Zhang, Boya Wu, Qicong Chen, Juanzi
ber of numerous international conferences in the areas of computer
Li, Chunxiao Xing, and Jie Tang. How do your friends on social
graphics, multimedia, and text processing. He was the Conference
media disclose your emotions? In Proceedings of the 28th AAAI
Co-Chair of ACM Multimedia in 2005, the Conference on Image and
Conference on Artificial Intelligence (AAAI’14), pages 306–312, 2011.
Video Retrieval in 2005 and ACM SIGIR in 2008, and the Technical
[50] Shuo Zeng, Mingfeng Lin, and Hsinchun Chen. Dynamic user- PC Co-Chair of SIGIR in 2010. He serves on the editorial boards of
level affect analysis in social media: Modeling violence in the dark the ACM Transactions of Information Systems, Foundation and Trends
web. In IEEE International Conference on Intelligence and Security in Information Retrieval, The Visual Computer, and Multimedia Tools
Informatics (ISI), pages 1–6, 2011. and Applications. He is on the Steering Committees of the International
[51] Qian Zhang and Bruno Goncalves. Topical differences between Conference on Multimedia Retrieval, Computer Graphics International,
chinese language twitter and sina weibo. Computer Science, 2015. and Multimedia Modeling Conference Series. He serves as a member
[52] Yuan Zhang, Jie Tang, Jimeng Sun, Yiran Chen, and Jinghai Rao. of international review panels of two large-scale research projects in
Moodcast: Emotion prediction via dynamic continuous factor Europe. He is the Independent Director of two listed companies in
graph model. 2013 IEEE 13th International Conference on Data Singapore.
Mining, pages 1193–1198, 2010.
[53] Jichang Zhao, Li Dong, Junjie Wu, and Ke Xu. Moodlens: an
emoticon-based sentiment analysis system for chinese tweets. In
Proceedings of the 18th ACM SIGKDD international conference on
Knowledge discovery and data mining, pages 1528–1531, 2012.

00 Intro
No ratings yet
00 Intro
18 pages
Detecting Stress Based On Social Interactions in Social Networks
No ratings yet
Detecting Stress Based On Social Interactions in Social Networks
2 pages
Deep Residual U Net For Automatic Detection OfMoroccan Coastal Upwelling Using SST Images
No ratings yet
Deep Residual U Net For Automatic Detection OfMoroccan Coastal Upwelling Using SST Images
5 pages
Final Doc Strees Detection
No ratings yet
Final Doc Strees Detection
57 pages
Detecting Stress Based On Social Interactions in Social Networks
100% (1)
Detecting Stress Based On Social Interactions in Social Networks
4 pages
Detecting Stress Based On Social Interactions in Social Networks
No ratings yet
Detecting Stress Based On Social Interactions in Social Networks
4 pages
Detecting Stress Based On Social Interactions in Social Networks
No ratings yet
Detecting Stress Based On Social Interactions in Social Networks
4 pages
Detecting Stress Based on Social Interactions in
No ratings yet
Detecting Stress Based on Social Interactions in
5 pages
PRC 1
No ratings yet
PRC 1
11 pages
2023 V14i40229
No ratings yet
2023 V14i40229
9 pages
Share 'Detecting Stress Based On Social Interactions in Social
100% (1)
Share 'Detecting Stress Based On Social Interactions in Social
11 pages
Synopsis
No ratings yet
Synopsis
12 pages
Project
No ratings yet
Project
13 pages
06 - Chapter 2
No ratings yet
06 - Chapter 2
22 pages
Paper - Psychological Stress Using Social Media
No ratings yet
Paper - Psychological Stress Using Social Media
12 pages
Teenagers' Stress Detection Based On Time-Sensitive Micro-Blog Comment/Response Actions
No ratings yet
Teenagers' Stress Detection Based On Time-Sensitive Micro-Blog Comment/Response Actions
11 pages
A Mental State Knowledge-Aware and Contrastive Network For Early
No ratings yet
A Mental State Knowledge-Aware and Contrastive Network For Early
16 pages
3223-Article Text-6272-1-10-20190531
No ratings yet
3223-Article Text-6272-1-10-20190531
12 pages
Calibration of Transformer-Based Models For
No ratings yet
Calibration of Transformer-Based Models For
12 pages
07 - Chapter 3
No ratings yet
07 - Chapter 3
6 pages
Beyond Depression Detection: Unveiling The Complexities of Social Media and Mental Health
No ratings yet
Beyond Depression Detection: Unveiling The Complexities of Social Media and Mental Health
9 pages
PRC 2
No ratings yet
PRC 2
15 pages
874
No ratings yet
874
6 pages
Less_is_More_Stress_Detection_through_Condensed_So
No ratings yet
Less_is_More_Stress_Detection_through_Condensed_So
10 pages
1 s2.0 S187705092202261X Main
No ratings yet
1 s2.0 S187705092202261X Main
10 pages
Sample Course End Project Report
No ratings yet
Sample Course End Project Report
25 pages
Research Paper (PREDICTION OF DEPRESSION LEVELS USING SOCIAL MEDIA)
No ratings yet
Research Paper (PREDICTION OF DEPRESSION LEVELS USING SOCIAL MEDIA)
11 pages
Machine Learning Based Mental Disorder Detection Through Emotions
No ratings yet
Machine Learning Based Mental Disorder Detection Through Emotions
24 pages
Abstract
No ratings yet
Abstract
21 pages
New Stress
No ratings yet
New Stress
52 pages
Detecting Depression and Mental Illness On Social Media
No ratings yet
Detecting Depression and Mental Illness On Social Media
7 pages
MISQ2020
No ratings yet
MISQ2020
25 pages
Data Sources
No ratings yet
Data Sources
141 pages
A-17 Paper
No ratings yet
A-17 Paper
4 pages
Methods in Predictive Techniques For Mental Health Status On Social Media: A Critical Review
No ratings yet
Methods in Predictive Techniques For Mental Health Status On Social Media: A Critical Review
11 pages
Depression PDF
No ratings yet
Depression PDF
12 pages
A Survey of Computational Methods for Online Mental State Assessment
No ratings yet
A Survey of Computational Methods for Online Mental State Assessment
34 pages
Research Paper
No ratings yet
Research Paper
2 pages
gao-et-al-2024-factor-structure-and-psychometric-properties-of-the-digital-stress-scale-in-a-chinese-college-sample
No ratings yet
gao-et-al-2024-factor-structure-and-psychometric-properties-of-the-digital-stress-scale-in-a-chinese-college-sample
13 pages
Forensis Review
No ratings yet
Forensis Review
19 pages
Base Paper
No ratings yet
Base Paper
7 pages
2019 A - Knowledge-Based - Recommendation - System - That - Includes - Sentiment - Analysis - and - Deep - Learning
No ratings yet
2019 A - Knowledge-Based - Recommendation - System - That - Includes - Sentiment - Analysis - and - Deep - Learning
12 pages
depression detection review-2
No ratings yet
depression detection review-2
19 pages
Detecting Mental Distresses Using Social Behavior Analysis in The Context of COVID-19: A Survey
No ratings yet
Detecting Mental Distresses Using Social Behavior Analysis in The Context of COVID-19: A Survey
30 pages
Ismail2019 Chapter SocialMediaAndPsychologicalDis
No ratings yet
Ismail2019 Chapter SocialMediaAndPsychologicalDis
23 pages
Article - Development and Initial Evaluation of A Multidimensional Digital Stress Scale
No ratings yet
Article - Development and Initial Evaluation of A Multidimensional Digital Stress Scale
13 pages
NEW ABS
No ratings yet
NEW ABS
1 page
Review 2
No ratings yet
Review 2
19 pages
Exploratory Analysis of Social Media Prior To A Suicide Attempt
No ratings yet
Exploratory Analysis of Social Media Prior To A Suicide Attempt
12 pages
s41598-020-68764-y
No ratings yet
s41598-020-68764-y
6 pages
Untitled
No ratings yet
Untitled
8 pages
1 s2.0 S1532046419300577 Main
No ratings yet
1 s2.0 S1532046419300577 Main
22 pages
Mental Health Analysis in Social Media Posts: A Survey: Muskan Garg
No ratings yet
Mental Health Analysis in Social Media Posts: A Survey: Muskan Garg
24 pages
Research Paper FF
No ratings yet
Research Paper FF
18 pages
detecting depression in social media
No ratings yet
detecting depression in social media
8 pages
Harnessing_the_Power_of_Hugging_Face_Transformers_for_Predicting_Mental_Health_Disorders_in_Social_Networks
No ratings yet
Harnessing_the_Power_of_Hugging_Face_Transformers_for_Predicting_Mental_Health_Disorders_in_Social_Networks
11 pages
sensors-22-09775-v2
No ratings yet
sensors-22-09775-v2
28 pages
Researching Mental Health Disorders in The Era of Social Media - Systematic Review
No ratings yet
Researching Mental Health Disorders in The Era of Social Media - Systematic Review
17 pages
Should I Post The Relationships Among Social Media Use, Emotion Recognition, and Mental Health
No ratings yet
Should I Post The Relationships Among Social Media Use, Emotion Recognition, and Mental Health
11 pages
Stressor Classification of Filipino Political Tweets Using Lda, SVM, Xgboost, Logistic Regression
No ratings yet
Stressor Classification of Filipino Political Tweets Using Lda, SVM, Xgboost, Logistic Regression
11 pages
Ipseology: A new science of the self
From Everand
Ipseology: A new science of the self
Dr. Jason Jeffrey Jones
No ratings yet
Netliving 101: Networking Life's Journey
From Everand
Netliving 101: Networking Life's Journey
John C. Durkin
No ratings yet
Road Assistance System Using GPS
No ratings yet
Road Assistance System Using GPS
2 pages
Research On Kano Model Based On Online Comment Data Mining: Huaming Song Chao Chen
No ratings yet
Research On Kano Model Based On Online Comment Data Mining: Huaming Song Chao Chen
7 pages
Rental House Web Portal
No ratings yet
Rental House Web Portal
2 pages
Road Assistance System Using GPS: ISSN: 2454-132X Impact Factor: 4.295
No ratings yet
Road Assistance System Using GPS: ISSN: 2454-132X Impact Factor: 4.295
7 pages
Research On Kano Model Based On Online Comment Data Mining
No ratings yet
Research On Kano Model Based On Online Comment Data Mining
2 pages
Online Faculty Staff Directory For Multiple University
No ratings yet
Online Faculty Staff Directory For Multiple University
2 pages
Online Bookstore - A New Trend in Textbook Sales Management For Services Marketing
No ratings yet
Online Bookstore - A New Trend in Textbook Sales Management For Services Marketing
14 pages
Privacy-Preserving Location-Proximity For Mobile Apps
No ratings yet
Privacy-Preserving Location-Proximity For Mobile Apps
2 pages
MLM - Multilevel Marketing System
No ratings yet
MLM - Multilevel Marketing System
1 page
Geo Trends Classification Over Maps Android
No ratings yet
Geo Trends Classification Over Maps Android
4 pages
My Forum Portal For Technical Help
No ratings yet
My Forum Portal For Technical Help
1 page
Fake Product Review Monitoring and Removal For Genuine Online Product Reviews Using Opinion Mining
No ratings yet
Fake Product Review Monitoring and Removal For Genuine Online Product Reviews Using Opinion Mining
4 pages
Location Based Information App For Ambulance and Police
No ratings yet
Location Based Information App For Ambulance and Police
1 page
Comments Mining With TF-IDF: The Inherent Bias and Its Removal
No ratings yet
Comments Mining With TF-IDF: The Inherent Bias and Its Removal
14 pages
College Networking Site Using PHP
No ratings yet
College Networking Site Using PHP
2 pages
ECommerce Postal For Electrical Product
No ratings yet
ECommerce Postal For Electrical Product
1 page
Catch You If You Misbehave
No ratings yet
Catch You If You Misbehave
3 pages
Cloud-Based Web Application With NFC For Employee Attendance Management System
No ratings yet
Cloud-Based Web Application With NFC For Employee Attendance Management System
6 pages
City Cab Booking and Reporting System
No ratings yet
City Cab Booking and Reporting System
1 page
Application of Text Classification and Clustering of Twitter Data For Business Analytics
No ratings yet
Application of Text Classification and Clustering of Twitter Data For Business Analytics
2 pages
Cloud-Based Web Application With NFC For Employee Attendance Management System
No ratings yet
Cloud-Based Web Application With NFC For Employee Attendance Management System
2 pages
P068
No ratings yet
P068
4 pages
VIVA-QUESTIONS-FOR-DS-LAB
No ratings yet
VIVA-QUESTIONS-FOR-DS-LAB
10 pages
1 s2.0 S1569843222000619 Main
No ratings yet
1 s2.0 S1569843222000619 Main
20 pages
Deeppose Estimation in Wild
No ratings yet
Deeppose Estimation in Wild
10 pages
AI Trainer (Mediapipe) (Futurescope) - 2023
No ratings yet
AI Trainer (Mediapipe) (Futurescope) - 2023
8 pages
Full Download Deep Learning in Bioinformatics: Techniques and Applications in Practice Habib Izadkhah PDF
100% (5)
Full Download Deep Learning in Bioinformatics: Techniques and Applications in Practice Habib Izadkhah PDF
64 pages
Toward More Accurate Heterogeneous Iris Recognition With Transformers and Capsules
No ratings yet
Toward More Accurate Heterogeneous Iris Recognition With Transformers and Capsules
13 pages
The Little Book of Deep Learning
100% (1)
The Little Book of Deep Learning
140 pages
DH GAN Mstar
No ratings yet
DH GAN Mstar
11 pages
Memristor Computing Systems Leon O. Chua (Editor) All Chapters Instant Download
100% (3)
Memristor Computing Systems Leon O. Chua (Editor) All Chapters Instant Download
37 pages
Potential Question Slot 3-4
No ratings yet
Potential Question Slot 3-4
12 pages
Complete Fuzzy Machine Learning Algorithms For Remote Sensing Image Classification 1st Edition Anil Kumar PDF For All Chapters
100% (2)
Complete Fuzzy Machine Learning Algorithms For Remote Sensing Image Classification 1st Edition Anil Kumar PDF For All Chapters
62 pages
Self-Supervised Deep Correlation Tracking
No ratings yet
Self-Supervised Deep Correlation Tracking
10 pages
Soranson Python-Machine-Learning RuLit Me 683600
No ratings yet
Soranson Python-Machine-Learning RuLit Me 683600
99 pages
LECTURE#9 EE258 F22 Part2 Draft v1
No ratings yet
LECTURE#9 EE258 F22 Part2 Draft v1
14 pages
Classification of ECG Signal Using FFT Based Impro
No ratings yet
Classification of ECG Signal Using FFT Based Impro
22 pages
Paper 2728
No ratings yet
Paper 2728
10 pages
Natural Disaster Prediction Model Using Deep Learning
No ratings yet
Natural Disaster Prediction Model Using Deep Learning
2 pages
Network Morphism
No ratings yet
Network Morphism
11 pages
Early Ginger Disease Detection Using Deep Learning Approach3
No ratings yet
Early Ginger Disease Detection Using Deep Learning Approach3
10 pages
aksdcbfd12324_4354665tgdfhghdf
No ratings yet
aksdcbfd12324_4354665tgdfhghdf
11 pages
Instance Segmentation For Autonomous Vehicle
No ratings yet
Instance Segmentation For Autonomous Vehicle
6 pages
Factored Systolic Array Tensor Processing
No ratings yet
Factored Systolic Array Tensor Processing
7 pages
Autonomous Drone Surveillance System
No ratings yet
Autonomous Drone Surveillance System
52 pages
Zefri Et Al - 2023 - Advanced Classification of Failure-Related Patterns On Solar Photovoltaic Farms
No ratings yet
Zefri Et Al - 2023 - Advanced Classification of Failure-Related Patterns On Solar Photovoltaic Farms
6 pages
SLICING AIDEDHYPERINFERENCEANDFINE-TUNING FORSMALLOBJECTDETECTION
No ratings yet
SLICING AIDEDHYPERINFERENCEANDFINE-TUNING FORSMALLOBJECTDETECTION
5 pages
Artificial Neural Networks: An Overview: August 2023
No ratings yet
Artificial Neural Networks: An Overview: August 2023
11 pages
Deep Learning Report
No ratings yet
Deep Learning Report
21 pages
Unit-2
No ratings yet
Unit-2
28 pages