helpfn

Predicting Social Emotions from Readers’ Perspective

Xintong Li , Qinke Peng , Zhi Sun, Ling Chai, and Ying Wang

Abstract—Due to the rapid development of Web, large numbers of documents assigned by readers’ emotions have been generated

through new portals. Comparing to the previous studies which focused on author’s perspective, our research focuses on readers’

emotions invoked by news articles. Our research provides meaningful assistance in social media application such as sentiment

retrieval, opinion summarization and election prediction. In this paper, we predict the readers’ emotion of news based on the social

opinion network. More specifically, we construct the opinion network based on the semantic distance. The communities in the news

network indicate specific events which are related to the emotions. Therefore, the opinion network serves as the lexicon between

events and corresponding emotions. We leverage neighbor relationship in network to predict readers’ emotions. As a result, our

methods obtain better result than the state-of-the-art methods. Moreover, we developed a growing strategy to prune the network

for practical application. The experiment verifies the rationality of the reduction for application.

Index Terms—Affect sensing and analysis, recognition of group emotion, affective text mining, complex network

Ç

1 INTRODUCTION

SOCIAL emotion prediction is of value to market analysisand to political decision [1], [2], [3], [4]. With the free and convenient communication environment of internet, people show increasing enthusiasm of online communication [5], [6]. Meanwhile, the internet users prefer to produce and convey online information through expressing personal opinions than just obtain online information. In this way, numerous news articles and comments have been published and shared rapidly via social media services. As a result, abundant under- lying positive or negative emotion information spreads and reflects the social sentiment tendency. Most intuitively, emo- tional label has been widely used in social web services. Fig. 1 indicates the result of voting for a news article using emotion labels from a popular news portal. Large numbers of people concerned about a hot news online. Therefore, valuable and available emotional information is continuously provided for scientific research work [7], [8], [9]. Furthermore, comparing to the traditional methods, which need to do numbers of sur- veys offline, data processing technology has been developed more feasible in the field of emotional extraction, analysis and prediction with its benefits of lower cost, higher efficiency and more accuracy. Under this circumstance, readers’ emo- tions prediction shows a highly research potential.

Compared with the typical tasks of sentiment analysis, opinion mining or affect recognition which based on

subjective text, social opinion prediction focuses on objec- tive text, for example news articles, which may not contain any opinion, but can evoke readers’ certain emotion. Due to the particularity of the task, social opinion prediction has potential applications which are different from those of writer-sentiment analysis [10]. Considering the effect of social media on the public sentiment, social emotion analy- sis engenders large benefits to social and economic problem, such as political issues and brand perception.

In this paper, we implement social opinion prediction by generating a real-time social opinion network. In more details, first, we train word vectors according to the most recent Wikipedia word corpus. Second, we calculate seman- tic distance between news via word vectors. As a metric between opinions, semantic distance allows us to construct the opinions growing network to describe the dynamical social opinions. Last, we predict follow-up news’ social emotion based on the network.

The opinion growing network naturally incorporates the knowledge encoded in the word2vec space and leads to high performance—the predicting via network outper- forms all 9 state-of-the-art emotion models [8], [11], [12], [13], [14], [15], [16], [17], [18], [19]. It is highly interpret- able based on the opinion growing network. The opinion growing network as stereotypical knowledge grows as the follow-up news join. With the enrichment of knowl- edge, the prediction results become more accurate. More- over, we can leverage existing networks to provide visual summaries to analyze social emotions evolution and social hotspots.

The rest of this article is organized as follows. In Section 2, we provide related work. In Section 3, we discuss methodol- ogy and key techniques in detail. In Section 4, we illustrate and discuss the dataset, experimental results. In Section 5, we provide concluding remark.

� The authors are with the Systems Engineering Institute, Xi’an Jiaotong University, Xi’an 710049, China. E-mail: {lixintong, sun.zhi, chailing} @stu.xjtu.edu.cn, {qkpeng, ying_wang}@xjtu.edu.cn.

Manuscript received 15 Sept. 2016; revised 8 Apr. 2017; accepted 16 Apr. 2017. Date of publication 18 Apr. 2017; date of current version 6 June 2019. (Corresponding author: Qinke Peng.) Recommended for acceptance by F. Sebastiani. For information on obtaining reprints of this article, please send e-mail to: [email protected], and reference the Digital Object Identifier below. Digital Object Identifier no. 10.1109/TAFFC.2017.2695607

IEEE TRANSACTIONS ON AFFECTIVE COMPUTING, VOL. 10, NO. 2, APRIL-JUNE 2019 255

1949-3045 � 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See ht _tp://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: University of the Cumberlands. Downloaded on September 25,2021 at 02:23:41 UTC from IEEE Xplore. Restrictions apply.

2 RELATED WORK

Social opinion prediction is a difficult research endeavor. As the initial research work on social opinion prediction, “affective text” in SemEval-2007 Tasks [11], [20] intend to annotate news headlines for the evoked emotion of readers. Another research focus on readers’ emotion evoked by news sentences [21]. Existing methods of social opinion prediction can be divided into three categories: knowledge-based tech- niques, statistical methods and hybrid approaches [3].

Because of the deficiency of information of news text [20], [22], it is unmanageable to annotate the emotions consis- tently. Knowledge-based techniques utilize existing emo- tional lexicon to supplement the prior knowledge for annotating the emotions. The popular emotional lexicon includes Affective Lexicon [23], linguistic annotation scheme [24], WordNet-Affect [25], SentiWordNet [26], and SenticNet [27], [28]. The drawback of knowledge-based techniques is the reliance on the coverage of the emotional lexicon. These techniques cannot process terms that do not appear in the emotional lexicon.

Statistical methods predict social opinion by training a sta- tistical model based on a large number of well-labeled corpus. There are two principal categories of statistical methods: word-level [11], [20], [29], [30], [31] and topic-level [7], [8], [12], [13], [14] methods. Word-level methods focus on exploit- ing the sentiment of individual words [11], [20] on the idea that words are the foundation of user sentiments. In order to model the word-emotion association, a variant of Na€ıve Bayes model named Emotion-Term(ET) is created [8]. The words extracted from the news articles are considered as indepen- dent features which indicate the emotion. However, word- level features in social opinion prediction are always inter- fered by the background noise words [7]. In particular, the methods treat each word individually, many emotional words are usually mixed with background noise words. In addition, the methods usually utilize the bag of words model to represent the text. In order to ensure the accuracy of emo- tional recognition, a large number of annotations are required. More recently, topic-level methods try to exploit the senti- ment of topics. A real-world event, object, or abstract entity that is the primary subject of the opinion as intended by the opinion holder can be regarded as a topic in the topic-level social opinion prediction [32]. The machinery of latent topic models like the Latent Dirichlet Allocation (LDA) is adopted in the Emotion-Topic Model (ETM) [8], [33]. The ETM [7], [8] added an intermediate layer into LDA so that a topic became vitally relevant to an emotion. According to the experiments on grouping topics into different emotions, ETM surpasses several other methods including SVM for social opinion pre- diction [7], [8]. Many homologous model to ETM were designed, such as Author-Topic Model(ATM) [34], Labeled LDA [35], [36] and Joint Sentiment/Topic Model (JSTM) [37].

Relative to LDA, ATM assumes that each author complies with a Dirichlet distribution over topics while each topic com- plies with a Multinomial distribution over words. Labeled LDA integrates correspondence between LDA’s latent topics and user tags. JSTM works well expressly for sentiment analy- sis of movie reviews. Those models above are all formed from the perspective of authors. Writers may have their favored topics or personal impression before writing an article or a review. It means that emotions can predefined and be showed in the latent topics and observable words. However, readers’ emotion ratings are not firmly determined until they read the content of a news article. Moreover, topic-level models are formed with parameters. It requires tuning strategy for the parameters, especially the number of topics. Statistical methods focus on data-driven modeling. However, senti- ment-related phenomena are complex human subjectivity. Topic-level methods employ a hidden assumption that there is an unknown author for each document which is deviant. This lack of explanatory leads to poor apprehen- sion and confusion about what concepts or features should be involved in the text analysis [38], [39]. At the same time, in the case of insufficient data volume, the methods are dif- ficult to obtain satisfactory results.

Hybrid approaches combine knowledge-based techni- ques and statistical methods. In short, the models utilize the prior knowledge as the text feature combined with statisti- cal methods for emotional prediction. Sentic [40] utilizes a priori knowledge to identify linguistic patterns combined with statistical models to predict emotions from text. Soujanya [18], [19], [41], [42] utilizes word2vec [43] to get word embeddings based on Google News corpus. Each arti- cle is represented by the word embedding in the order of the words in the article. The vector of the article act as Con- volutional Neural Network’s (CNN) input for training. The trained CNN can be utilized as a trainable feature extractor. The internal output of CNN is taken as the corresponding text feature. Finally, Support Vector Machine (SVM) classi- fier is trained and achieves satisfactory prediction result.

3 PROBLEM DEFINITION AND ANALYSIS

In this section, the problem about social opinion prediction is well defined, including the relevant general terms and notations. Later on, we describe the social opinion model and opinion network in detail. Finally, we propose a strat- egy for social opinion prediction.

3.1 Problem Formulation

We define the following notations for describing the social opinion prediction: An online news collection D consists of news d, and the emotion ratings labels E. The list of emotion labels is denoted by E¼fekg, and ek indicates emotion titled “joy”, “anger”, “fear”, “surprise”, “touching”, “empathy”, “boredom”, “sadness”, “warmness” etc. In particular, a news d is a set of word tokens W¼fwig, and a set of ratings over E emotion labels denoted by vd ¼ fvd;eg. The value of vd;ek is the number of online users who have voted the kth emotion label ek for news d. Table 1 summarizes the nota- tions of these frequently-used variables.

According to Kim and Hovy [44], the opinion can be split into four parts: topic, opinion holder, claim, and sentiment.

Fig. 1. An example of emotion labels and user ratings.

256 IEEE TRANSACTIONS ON AFFECTIVE COMPUTING, VOL. 10, NO. 2, APRIL-JUNE 2019

Authorized licensed use limited to: University of the Cumberlands. Downloaded on September 25,2021 at 02:23:41 UTC from IEEE Xplore. Restrictions apply.

To be specific, a holder believes a claim about a topic with a sentiment. For social opinion, the opinion holder stands for users who have voted the news. The topic can be replaced by the content of the news. The sentiment can be measured by the vote around the set of predefined emotional labels. The claim is unobservable and inessential in this task. This kind of social opinion can be model with a quadruple hevent; f; s; ti, where event stands for the social event; f is the text feature set of social event; s is the result of voting towards social event which is represented as distribution over the predefined emotional labels. t is the time when the social events occurred. The social opinion prediction task this paper discussed is focused on the prediction of s based on the former social opinion quadruples.

3.2 Social Opinion Model

Sentiment-related phenomena can be explained as the process of evaluation of events, objects or persons [23], [45], [46]. The opinions are caused by the subjective evaluation of the “raw” stimuli. The “raw” stimuli may have no intrinsic emotional meaning, but will be appraised by personal relevance and implications [47]. For social opinion, the “raw” stimuli are only text and its features which are difficult to expound the corresponding sentiment-related phenomena. In fact, there are less than 5 percent of directly emotional words of a text in daily speech, emotional writing, and affect-laden poetry [48]. In journalism domain, a lower percentage is undisputed. It is rarely influenced by the personal relevance under the social community. To simplify the problem, we focus on implica- tions without personal relevance.

According to the cognitive approaches, the result of vot- ing is “the person’s experience, goals and opportunities for action” [43]. It is process that evaluates an event by dimen- sions such as urgency, consistency with goals, etc. All the social opinions share the similar emotional experience, goals and opportunities for action with each other. From the NLP perspective, the models are inexplicable but feasible. From psychology and linguistics perspective, the models are explicable but lack of use in the service. Based on the general characteristic, similarity is one of six principles that guide human perception of the world in Gestalt theory [28],

[49]. We can predict social opinions by measuring the semantic similarity between events.

The social cognitive process can be modeled based on a stereotypical knowledge set consisting of social opinion P¼fhevent; f; s; ti1; hevent; f; s; ti2; . . . ; hevent; f; s; tiNg. Instead of establishing appraisal criteria, the cognitive process can be regarded as the neighbor analysis in set P. The set P can be interpreted as social experience. The cognitive process can be simplified as matching the “raw” stimuli between the priori social opinions in set P. The social community has sta- bilized emotion towards specific events. It can be explained by a social psychology behavior named “stereotype” which is a fixed view of people, groups, events, institutions, or problems [50], [51], [52]. Stereotype widely exists in media [50], [51]. To be more accurate, the task is modeling the rela- tionship between current event and priori social opinions based on semantic similarity.

As we defined, the social opinion is a quadruple hevent; f;s; ti. In social opinion model, the first step is called words- extract in which the raw feature (f) of social event is processed as Bag of Words (BOW) which is widely used for document representation. Considering the online performance of the algorithm, we utilize the term frequency(TF) instead of term frequency–inverse document frequency (TF-IDF) for measur- ing the importance of a word in corpus. TF-IDF is not avail- able in real-time updated data. Because we need to know the global distribution of words to calculate TF-IDF. By contrast, TF can be explained as social community ratio of attention to the words. Based on the explanation of attention, it is persua- sive that the readers concern each news equally, in other words, pay equal attention to each news. Based on it, the assumption is justifiable that the BOW f should be normal- ized. So far, the preprocessed fi in social opinion quadruple hevent; f;s; tii is a normalized histogram which is finite- dimensional vectors with nonnegative coordinates whose sum is equal to 1. That is fi 2 Sw Sw ¼D fu 2 Rwþj

Pw i¼1 ui ¼ 1g.

We denote fij as the weight of jth word in ith event. Considering the implications in social opinion feature f,

we leverage the recent result named word2vec [53] which shows that a log-bilinear model can learn high quality embedding of words by local co-occurrences in sentences. An embedding of a word is a finite-dimensional vector which express the word’s meaning. We can measure the semantic distance between two words by Euclidean dis- tance. The stable word-embedding is a cornerstone of our model. We require large corpora to perform word2vec. The learning corpora can influence the quality of embedding directly. It is reasonable to choose Wikipedia as learning corpora. Comparing with the web-based text collected from online media, Wikipedia is a free online encyclopedia con- sisting of various entries which ensures completeness of words so then stability of the model. Moreover, it ensures that the model can provide strong robustness because of Wikipedia’s real-time self-renewal.

The Euclidean distance between words in word2vec space measures the semantic similarity. More precisely, cði; jÞ ¼ kxi � xjk2 denotes the distance between word i and word j where xi and xj represent corresponding word. By measuring the similarity between social opinions hevent; f; s; ti1 and hevent; f; s; ti2, can we model the rela- tionship. As f is a normalized histogram and each word in

TABLE 1 Notations of Frequently-Used Variables

Notations Description

D All the news wd;k The kth word token in news d W ¼ fwig The set of vocabulary ek The kth emotion label E ¼ fekg The list of emotion labels the common

instances of E are “joy”, “anger”, “fear”, “surprise”, “touching”, “empathy”, “boredom”, “sadness”, “warmness”, etc.

vd ¼ fvd;ekg The set of ratings over E emotion labels vd; ek

the number of online users who have voted the kth emotion label ek for news d

v̂n The prediction of emotion conditioned on future unlabeled news d which only contain words.

si;j The similarity between news i and news j cði; jÞ The distance between word wi and word wj xi The vector representation for word wi

LI ET AL.: PREDICTING SOCIAL EMOTIONS FROM READERS’ PERSPECTIVE 257

Authorized licensed use limited to: University of the Cumberlands. Downloaded on September 25,2021 at 02:23:41 UTC from IEEE Xplore. Restrictions apply.

f1 has a meaning distance to each word in f2. To compare two histograms f1 and f2, we apply optimal transport dis- tance. The task of optimal transport distance between two histograms can be illustrated in Fig. 2. Word1-4 in f1 can be moved to word5-7 in f2 at a cost cði; jÞ based on the seman- tic distance. The semantic transport task between f1 and f2 is to move all the words in f1 to f2.

We utilize the word mover’s distance metric [50] as our distance. The measurement of similarity can therefore be achieved, in principle, by the measurement of optimal trans- port distance. First, we assume that word i in f1 can be transformed totally or partially into any word in f2. There exists a flow matrix T 2 <n�n, in which element Tij � 0 indi- cates the extent of switch from word i in f1 to word j in f2. The conditions of transforming f1 entirely into f2 are that:

(1) the amount of outgoing flow from word i equals f1i, i.e.

P j Tij ¼ f1i;

(2) likewise, the sum of incoming flow to word j equals f2j, i.e.

P i Tij ¼f2j.

Now, the distance between the two histograms are defined as the minimum (weighted) cumulative cost required to transform all words from f1 to f2, i.e.P

i;j Tijcði; jÞ. It can be described as following:

min T�0

Xn i;j¼1

Tijc i; jð Þ

Subject to : Xn j¼1

Tij ¼ f1i 8i 2 1; � � �; nf g

Xn i¼1

Tij ¼ f2j 8j 2 1; � � �; nf g:

(1)

By solving the linear program above, can we model the relation between opinions as opinion distance.

3.3 Opinion Network

Based on the relation between opinions, we construct an opinion network, in which nodes indicate opinions and edges indicate relation between opinions. The opinion net- work acts as stereotypical knowledge which can serve as the lexicon between events and corresponding emotions. We can predict the social opinion through the network.

To construct the opinion network, we add the edge between nodes to denote distance. Fig. 3 presents the distri- bution of opinions distance from a real-world data. The figure shows that the distribution of opinions distance obeys Gaussian distribution. Then we explore the relation- ship between network structure and social emotions. Since the opinion network here is fully connected networks, we filter the edge to visualize the network structure. We prune edges shorter than threshold and label the nodes of 8 emo- tions by color. The network structure is different in thresh- olds. Here we choose the threshold for the visualization of the network. The threshold is chosen as 0.7 manually for visualization of network. We utilize ForceAtlas2 [54] algo- rithm to arrange the layout of nodes. The network shows in Fig. 4. The color of the nodes denotes the most voted emo- tion label in each news. The weights of edges reflect the value of distance.

There are emotion clusters in network through the Fig. 4. Then we unfold communities in the opinion network. The community structure of network is a set of highly inter- connected nodes [55] corresponding with the event in social opinion quadruple. The communities in network is shown in Fig. 5 in same color. Each community is regarded as an individual event containing news related to each other in social opinion quadruple hevent; f; s; ti. It is clear that event stands for a specific event and its follow-up story. We can

Fig. 2. Optimal transport distance between two histograms. Fig. 3. Distribution of similarity.

Fig. 4. Emotion in opinions network.

258 IEEE TRANSACTIONS ON AFFECTIVE COMPUTING, VOL. 10, NO. 2, APRIL-JUNE 2019

Authorized licensed use limited to: University of the Cumberlands. Downloaded on September 25,2021 at 02:23:41 UTC from IEEE Xplore. Restrictions apply.

control the granularity of event by regulating the threshold of the edge. According to the result of unfolding communi- ties, the event stands for a class of events like medical dis- pute etc.

Compared the community structure of network with the emotion distribution in network, there are correlations between emotion and network architecture based on the Figs. 4 and 5. The social community has stabilized emotion towards specific events. In other words, all elements in the same community (event) of opinions network hold the simi- lar emotional voting result s. To evaluate the similarity of emotional voting result in the same community for testing our inference, we present evaluation method Community Mean Pearson’s correlation coefficient (CMP) and Commu- nity Variance Pearson’s correlation coefficient (CVP). Con- sidering evaluating similarity between news emotional rating vi; vj, it is sensible to utilize Pearson’s correlation coefficient. The Pearson’s correlation coefficient between emotional rating vi; vj is evaluated as follow:

ri;j ¼ Cor vi; vj � �

¼ Cov vi; vj � �

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var við ÞVar vj

� �q : (2) To evaluate the regularity of a community, the Pearson’s

correlation coefficient is extended to the community struc- ture of network. CMP evaluates the average correlation level of a community which ranges (�1, 1) as same as the Pearson’s correlation coefficient. The CMP is ameliorated by the following:

CMP ¼ P

i;j2C;i 6¼j ri;j Cj j2 � Cj j

: (3)

CMP can only evaluate the overall correlation. To evalu- ate the dispersion of correlation in a community, CVP is ameliorated as follow:

CVP ¼ P

i;j2C;i 6¼j ri;j � CMP � �2

Cj j2 � Cj j : (4)

The lower CVP is, the smoother correlation exists in a community.

With the help of CMP and CVP, we can take a quantita- tive analysis about the correlation between opinions

emotional voting result s and event in opinions network. Besides CMP and CVP, the mean number of voting in com- munity (MN) is a metric which represents the reliability of the raw emotional voting result. The more voting, the more credible that voting result can mirror the actual social opinions.

Fig. 6 shows the CMP, CVP and MN distribution over communities. There are 34 distinct communities in opinions network corresponding to 34 events. The high value of CMP expresses similarity of emotional rating in community. Result shows that most events get high CMP with small CVP ensuring the stability. For poor performance result like small CMP with large CVP, they all map to low MN without exception. MN is the reliability of the raw emotional voting result. It is feasible that a high MN guarantees the precision of raw data mirroring social opinions. The rationality of result guarantees the accuracy of emotion prediction we proposed below.

Opinions network is a growing network. The news data is a time sequence of news documents. We utilize follow-up news to expand the network size. As time grows, news data continues to enrich the network. As the size of the network becomes larger, the storage and computing complexity grow, and eventually, the network becomes difficult to han- dle. In addition, the social opinion will change over time. Social emotions for a particular event are not immutable. In order to make up for the two drawbacks above, the out- dated inaccurate opinion node in the network should be removed. We propose a simple threshold-based network growing strategy for pruning the network. Fig. 7 shows opinion network growing strategy: At time T, a news joins the network as a green node k, which is close to the existing blue node i. While the distance between nodes is less than the specific threshold. So, at time T the node k replaces the node i for pruning. The growing strategy will influence the

Fig. 5. Communities in opinions network.

Fig. 6. Evaluation distribution in community.

Fig. 7. Network growing strategy.

LI ET AL.: PREDICTING SOCIAL EMOTIONS FROM READERS’ PERSPECTIVE 259

Authorized licensed use limited to: University of the Cumberlands. Downloaded on September 25,2021 at 02:23:41 UTC from IEEE Xplore. Restrictions apply.

results of emotion prediction, while the threshold should be selected according to the actual situation. The impact of net- work size will be discussed in the experimental part.

3.4 Social Opinion Prediction

Although there is the correlation between emotion and com- munity in opinions network, it cannot be utilized directly for its limitations. On the one hand, the threshold for prun- ing edges depends on the current network state. It is unable to determine the appropriate threshold of filter for the com- munity structure. On the other hand, the high temporal complexity of community structure partitioning algorithm is not conducive to online application. To predict emotion conditioned on future unlabeled opinion di which only con- tain words. We leverage the inference of community struc- ture of opinions network and simplify it as neighbor analysis. So that, we can predict the social opinion based on the opinions growing network.

First, in order to convert the distance into a standardized similarity, we are inspired by the literature [56] that trans- forms the uncertain distance relation of the opinion distance into a conditional probability to express the similarity. In more detail, we consider the two nodes i, j in the network. The node i chooses node j as the approaching node with the probability pi;j, considering the Gaussian distribution with node i as the center point. If j is closer to i, then pi;j is larger. And vice versa, if the distance is far, pi;j is extremely small. So, we can define pi;j as follows:

pi;j ¼ exp �d2i;j=2s2

� � P

k 6¼i exp �d2i;k=2s2 � � : (5)

Since we only care about the similarity of different points, we set pi;i ¼ 0.

Based on the nearest neighbor analysis. The bvi is esti- mated by the following:

S ¼ v0; � � �; vi�1f g v̂i ¼

X j2Stop

pi;jvj; (6)

where Stop contains the k-top nearest nodes with node i. There are two parameters that need to be determined: the upper limit number k of approaching node, s represents the normal distribution variance. They can be adjusted on train data. In this paper, we utilize a simple parameter optimization strat- egy: first set k as the fixed value, search the optimal s in the training set; and then search for the optimal k of the prediction effect under the current optimal parameter s; repeat iterations until k no longer changes. k and s always represent the attrib- utes of the current network. As the size of the network grows, the parameters are constantly adjusted. The prediction model we name it as Social Opinion Mining model.

4 EXPERIMENTS AND ANALYSIS

In this section, we present the experimental result and eval- uate the performance of the proposed models for social opinion prediction, then compare it with state of the art models. Moreover, we design the experiments to analyze the impact of network size on social opinion prediction.

4.1 Experiment Setup

To test the effectiveness of the proposed model, we utilize two datasets for testing (The dataset is available in public: github.com/lixintong1992/SocialEmotionData). One dataset used here is Yanghui Rao’s corpus [13] which collected 4570 news articles from the Society channel of Sina. The attributes of each article include the URL address, publishing date (from January to April of 2012), news title, content, and user ratings over 8 emotion labels: “touching”, “empathy”, “boredom”, “anger”, “amusement”, “sadness”, “surprise” and “warmness”. The other dataset is collected from the Soci- ety channel of Sina from January to December of 2016, a total of 5,258 hot news data. User ratings over 6 emotion labels: “touching”, “anger”, “amusement”, “sadness”, “surprise” and “curiosity”.

The average votes per news represent the number of votes for label. The average votes in dataset 2016 is 770.41 which means that the labels in each news are generated by approxi- mate 770 readers on average. By contrast, the average votes in dataset 2012 is 71.21, which means that the labels in each news are generated by only 71 readers on average. The label generated by more reader seems more credible. In addition, Fig. 8 shows the overall distance distribution between news. Both of the datasets obey normal distribution. But the expec- tation and variance of the normal distribution of dataset 2016 are greater than dataset 2012, which indicates the semantic correlation is relatively small. In summary, compared to dataset 2012, dataset 2016 has large time interval, less seman- tic correlation, more credible label.

We split the datasets into training and testing sets by chronological order. Because adjacent news articles can have a similar context, random division is unreasonable. For dataset 2012, 2,342 news published from January to Feb- ruary are used to construct opinions network and optimize parameters k and s. 2,228 news published from March to April are used for testing. For dataset 2016, 3,109 news pub- lished from January to June are used for training. 2,149 news published from July to November are used for testing. Table 2 shows the statistics for both datasets.

Before the modeling process, all the texts are transformed into histograms by vectorization and normalization. As a pre-training step, the corpus which contains only 4,570/ 5,258 news articles is insufficient to train the word-

Fig. 8. Distance distribution in both datasets.

260 IEEE TRANSACTIONS ON AFFECTIVE COMPUTING, VOL. 10, NO. 2, APRIL-JUNE 2019

Authorized licensed use limited to: University of the Cumberlands. Downloaded on September 25,2021 at 02:23:41 UTC from IEEE Xplore. Restrictions apply.

embedding. It’s irrational to learn the word representation in test data. We choose Wikipedia as training corpus for word representation by word2vec as we mentioned. The word-embedding trained by Wikipedia can be regarded as a lexicon. The lexicon consists of common terms. According to it, we can filter the text data by word-embedding lexicon.

4.2 Social Emotion Prediction

We conduct the prediction experiment to compare with the state of the art methods [13], [18], [19]. Based on the training data, we construct the online news growing network. For the testing data, we add the news node into network chronologi- cally. Then, we predict the news emotion by our model.

First, we evaluate the models presented above and dis- cuss the properties of models. In order to ensure the ratio- nality of the experimental results, we utilize the training data for network growing only without pruning. In predic- tion process, we keep the network constant to ensure the fairness of the results. The parameters k and s are constant for the optimal value of the training network. The parame- ters do not change during the prediction process. Then, to conduct a comprehensive comparison with various models, we compare the existing supervised unigram model (SWAT) [9], [12], emotion-term (ET), emotion-topic model (ETM) [5], [6], affective topic model (ATM) [16], multilabel supervised topic model (MSTM), sentiment latent topic model (SLTM) [15], Contextual Sentiment Topic Model (CSTM) [14] and deep learning approaches Convolutional Neural Network (CNN) [18], [19], [42], CNN-SVM [15].

For the problem of parameter tuning in topic-level mod- els, we experimented separately under different topic num- bers (2-30) according to the literature [12], and averaged the result as the final result.

For CNN and CNN-SVM, we use the CNN network struc- ture in literature [19]. For the input vectors in CNN, we create a concatenation vector of three parts:1. Left padding; 2. Text; 3. Right padding. Each of our articles uses 200 words for input. Each word is represented as 100 dimensions. So, the dimension of the input vector is 100� ð2 þ 200 þ 2Þ¼20;400. The CNN network structure is 7 layers:1. Input layer of 20,400 neurons; 2. Convolution layer, with a kernel size of 3 and 18 feature maps. The output of this layer was computed with a non-linear function, the hyperbolic tangent; 3. Max-pool layer with max-pool size of 2; 4. Convolution layer with kernel size of 4, 300 feature maps, also using the hyperbolic tangent; 5. Max-pool layer with max-pool size of 2; 6. Fully connected layer of 500 neutrons, whose values were later used as the

extracted features; 7. Output SoftMax layer of 6/8 neurons, depending on the number of labels.

For CNN model, we directly use the neural network out- put as the classification result. For CNN-SVM, we use CNN network as feature extractor. The output of the penultimate layer of the network (500-dimension vector) is the vector feature we use for training the SVM classifier.

We employ two evaluation metrics as indicators of per- formance: Acc@1 and the averaged Pearson’s correlation coefficient over all documents (AP).

The coarse-grained metric Acc@1 stands for the accuracy at top 1. According to Acc@1, a predicted ranked list of emo- tion labels is correct if the list’s first item is identical to the actual ranked list’s first item. If two emotion labels in the actual ranked list have the same number of votes, then their positions are interchangeable. Acc@1 is computed by divid- ing the number of correctly predicted documents by the total number of documents. The actual emotion set Etop includes the top-ranked emotion, and ep denotes the top- ranked predicted emotion. In this evaluation, the problem is formalized as a multi-class classification

Accd@1 ¼ 1 ep ¼ Etop0 else: �

(7)

Acc@1 ¼ P

d2D Accd@1 Dj j : (8)

Table 3 shows the Acc@1 of different models. The fine-grained metric AP is conducted because Acc@1

could not take emotional distributions into account. It ignores emotional distributions which is significant in social emotion mining. People’s emotions for a news are usually not very clear but a mixture of concentrated emotions. For each document, AP measures the correlation between the predicted distributions and the actual votes over all emotion labels. The value of AP ranges from �1 to 1, where 1 indi- cates a perfect positive correlation.

Cor vi; vj � �

¼ Cov vi; vj � �

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var við ÞVar vj

� �q (9)

AP ¼ P

i2D Cor vi; v̂ið Þ Dj j : (10)

Table 4 shows the AP of different models.

TABLE 2 Statistics of Data Sets

Emotion index 2012 2016

Training Testing Training Jesting

1 20,850 20,946 313,998 175,029 2 10,731 12,499 - - 3 10,519 11,476 - - 4 56,844 81,323 835,644 583,833 5 20,445 23,267 613,334 313,571 6 15,829 21,333 331,450 230,054 7 5,903 5,483 133,995 66,702 8 4,531 3,455 292,523 160,695

TABLE 3 Acc@1 Result to Different Models

Models Aec@l(%)

Dataset 2012 Dataset 2016

Social Opinion Mining model 61.27 58.59

Contextual sentiment topic model (CSTM) 54.95 40.74

Affective topic model (ATM) 49.58 29.20

Sentiment latent topic model (SLTM) 49.50 28.95

Multilabel supervised topic model (MSTM) 49.83 32.17

Emotion-topic model (ETM) 52.90 54.19

Supervised unigram model (SWAT) 50.63 38.97

Emotion-term (ET) 42.91 48.04

CNN 53.68 51.23

CNN-SVM 55.21 52.63

LI ET AL.: PREDICTING SOCIAL EMOTIONS FROM READERS’ PERSPECTIVE 261

Authorized licensed use limited to: University of the Cumberlands. Downloaded on September 25,2021 at 02:23:41 UTC from IEEE Xplore. Restrictions apply.

Compared to the other models, the Social Opinion Min- ing model outperforms others in both Acc@1 and AP metric. By contrast, the models perform better on dataset 2012. The main reason is that, as we mentioned, the time interval of dataset 2012 is shorter under the same amount of data. Meanwhile, dataset 2012’s semantic correlation between news is relatively higher. As a result, the prediction in data- set 2012 is better. For dataset 2016, we collect the daily hot news to ensure the quality of labels. So, the news covers a long period of time (about 1 year), but the semantic correla- tion between news is weak. The results show that for the dataset 2016, most models have different degrees of perfor- mance attenuation.

Our model is constructed based on the opinions network. Compared with other topic-level and word-level models, our model treats the news content and emotion distributing as a whole opinion structure. Word-level models simply separate the word and emotion respectively. It omits the correlation between words and emotions. Topic-level mod- els are generative models which assume that topics gener- ates emotions. However, topic-level models depend on the semantic correlation of dataset. The results show that CSTM, ATM, SLTM, MSTM, ETM perform well in the 2012 dataset. But their performance of the 2016 dataset is poor, only ETM maintains a good result. CSTM, ATM, SLTM, MSTM are relatively high complicated. The decline in the semantic correlation of data will lead to the decline in generalization performance. Our model is based on princi- ple of news content and emotion distribution jointly. In this way, we can retain the latent information between emotion distribution. As a result, ours outperform others both in coarse-grained and fine-grained metric.

4.3 Discussion

The results above do not take into account the network reduc- tion. In practical applications, maintenance of opinion net- work storage and calculation will be costly with the increasing number of nodes. Moreover, the social opinion towards specific news will change over time. Outdated news will affect the prediction results. More precisely, we do not need to remember the news 10 years ago. As we mentioned, the network growing strategy can control the size of the net- work by threshold. It makes the network more suitable in practice. We performed a battery of experiments to analyze the predicted results of the growing network after reduction.

To test the prediction results based on the reduced net- work, we utilize the reduction rules in the process of

generating the network to reduce the size of the network. The same testing set is used on the reduced network. The experimental results are shown in Fig. 9. The horizontal axis is the remaining proportion of the training set. The vertical axis is the result of the prediction. The dashed line is the best predictor of other models in the previous experiment based on the complete training set. The results show that for the dataset 2012, more than 16 percent of the training data can outperform the other results, for the dataset 2016, more than 40 percent of the training data can outperform the other results. As we mentioned, the different perfor- mance between the two datasets reflects the difference in the semantic correlation between news.

The final results show that the reduction strategy can effectively reduce the size of the network, while ensuring prediction accuracy. By adjusting the threshold, the net- work can adapt to different time interval, different actual requirements. It increases the versatility of the model.

5 CONCLUSION

In this paper, we analyze the online social opinions and pro- pose social opinion model for measuring similarity among news. Due to word-embedding pre-trained on Wikipedia, model’s stability and robustness are guaranteed and can hardly be influenced by the size of news data. Based on the

TABLE 4 AP Result to Different Models

Models AP

Dataset 2012 Dataset 2016

Social Opinion Mining model 0.62 0.64

Contextual sentiment topic model (CSTM) 0.52 0.43

Affective topic model (ATM) 0.44 0.28

Sentiment latent topic model (SLTM) 0.46 0.26

Multilabel supervised topic model (MSTM) 0.47 0.34

Emotion-topic model (ETM) 0.48 0.49

Supervised unigram model (SWAT) 0.47 0.40

Emotion-term (ET) 0.43 0.43

Fig. 9. Results based on the reduced network.

262 IEEE TRANSACTIONS ON AFFECTIVE COMPUTING, VOL. 10, NO. 2, APRIL-JUNE 2019

Authorized licensed use limited to: University of the Cumberlands. Downloaded on September 25,2021 at 02:23:41 UTC from IEEE Xplore. Restrictions apply.

similarity, we construct an opinion network to detect user- generated social emotion by the structures of opinion net- work. There is significant correlation between emotion and structures of news network as we expected. The perfor- mance of the prediction based on opinion network is more stable and accurate than existing models. In addition, we propose a threshold-based network growing strategy for pruning the network. The experiment verifies the rationality of the reduction: remaining 14 percent of the data can get the best results in interrelated dataset. For less relevant dataset, more than 40 percent of the data can guarantee the best results.

ACKNOWLEDGMENTS

We would like to thank the anonymous reviewers and edi- tor for their constructive comments and suggestions on ear- lier version of this paper. And we would like to thank Yanghui Rao for his helpful feedback. The work described in this paper has been supported by the National Natural Science Foundation of China (grant numbers 61173111).

REFERENCES [1] E. Cambria, B. Schuller, Y. Xia, and C. Havasi, “New avenues in

opinion mining and Sentiment analysis,” IEEE Intell. Syst., vol. 28, no. 2, pp. 15–21, Mar. 2013.

[2] E. Cambria, B. Schuller, Y. Xia, and B. White, “New avenues in knowledge bases for natural language processing,” Knowl.-Based Syst., vol. 108, pp. 1–4, Sep. 2016.

[3] E. Cambria, “Affective computing and Sentiment analysis,” IEEE Intell. Syst., vol. 31, no. 2, pp. 102–107, Mar. 2016.

[4] E. Cambria, N. Howard, Y. Xia, and T.-S. Chua, “Computational intelligence for big social data analysis [Guest Editorial],” IEEE Comput. Intell. Mag., vol. 11, no. 3, pp. 8–9, Aug. 2016.

[5] B. Zhang, X. Guan, M. J. Khan, and Y. Zhou, “A time-varying propagation model of hot topic on BBS sites and Blog networks,” Inf. Sci. (Ny)., vol. 187, pp. 15–32, 2012.

[6] Z. Sun, Q. Peng, J. Lv, and J. Zhang, “A prediction model of post subjects based on information lifecycle in forum,” Inf. Sci. (Ny)., vol. 337, pp. 59–71, 2016.

[7] S. Bao, et al., “Joint emotion-topic modeling for social affective text mining,” in Proc. IEEE Int. Conf. Data Mining, 2009, pp. 699–704.

[8] S. Bao, et al., “Mining social emotions from affective text,” IEEE Trans. Knowl. Data Eng., vol. 24, no. 9, pp. 1658–1670, Sep. 2012.

[9] Q. Wang, O. Wu, W. Hu, J. Yang, and W. Li, “Ranking social emo- tions by learning listwise preference,” in Proc. 1st Asian Conf. Pat- tern Recognit., 2011, pp. 164–168.

[10] K. H.-Y. Lin and H.-H. Chen, “Ranking reader emotions using pairwise loss minimization and emotional distribution regression,” in Proc. Conf. Empir. Methods Nat. Lang. Process., Oct. 2008, pp. 136–144.

[11] P. Katz, M. Singleton, and R. Wicentowski, “SWAT-MP: The SemEval-2007 Systems for Task 5 and Task 14,” in Proc. 4th Int. Workshop Semantic Eval., 2007, pp. 308–313.

[12] Y. Rao, “Contextual Sentiment topic model for adaptive social emotion classification,” IEEE Intell. Syst., vol. 31, no. 1, pp. 41–47, Jan. 2016.

[13] Y. Rao, Q. Li, X. Mao, and L. Wenyin, “Sentiment topic models for social emotion mining,” Inf. Sci. (Ny)., vol. 266, pp. 90–100, May 2014.

[14] Y. Rao, Q. Li, L. Wenyin, Q. Wu, and X. Quan, “Affective topic model for social emotion detection,” Neural Netw., vol. 58, no. 2012, pp. 29–37, Oct. 2014.

[15] S. Poria, E. Cambria, D. Hazarika, and P. Vij, “A deeper look into sarcastic Tweets using deep convolutional neural networks,” Col- ing, vol. 2016, pp. 1601–1612, Oct. 2016.

[16] S. Aral and D. Walker, “Identifying social influence in networks using randomized experiments,” IEEE Intell. Syst., vol. 26, no. 5, pp. 91–96, 2011.

[17] X. Li, J. Ouyang, and X. Zhou, “Supervised topic models for multi- label classification,” Neurocomputing, vol. 149, no. PB, pp. 811–819, 2015.

[18] Y. Kim, “Convolutional neural networks for sentence classi- fication,” in Proc. Conf. Empirical Methods Natural Language Process., vol. 21, no. 9, pp. 1746–1751, 2014.

[19] S. Poria, E. Cambria, and A. Gelbukh, “Deep convolutional neural network textual features and multiple kernel learning for utter- ance-level multimodal sentiment analysis,” in Proc. Conf. Empirical Methods Natural Language Process., Sep. 2015, pp. 2539–2544.

[20] C. Strapparava and R. Mihalcea, “Semeval-2007 task 14: Affective text,” in Proc. SemEval-2007, Jun. 2007, pp. 70–74.

[21] P. K. Bhowmick, “Reader perspective emotion analysis in text through ensemble based multi-label classification framework,” Comput. Intell. Secur., vol. 2, no. 4, pp. 64–74, 2009.

[22] C. Quan and F. Ren, “An exploration of features for recognizing word emotion,” in Proc. Int. Conf. Comput. Linguistics, 2010, pp. 922–930.

[23] A. Ortony, G. L. Clore, and A. Collins, “The cognitive structure of emotions,” Contemporary Sociology, vol. 18, no. 6, pp. 957–958, Nov. 1989.

[24] J. Wiebe, T. Wilson, and C. Cardie, “Annotating expressions of opinions and emotions in language,” Language Resources Eval., vol. 39, no. 2–3, pp. 165–210, 2005.

[25] C. Strapparava and A. Valitutti, “WordNet-Affect: An affective extension of WordNet,” in Proc. 4th Int. Conf. Lang. Resour. Eval., 2004, pp. 1083–1086.

[26] A. Esuli and F. Sebastiani, “SENTIWORDNET: A publicly avail- able lexical resource for opinion mining,” in Proc. 5th Conf. Lang. Resour. Eval., 2006, pp. 417–422.

[27] E. Cambria, D. Olsher, and D. Rajagopal, “SenticNet 3: A common and common-sense knowledge base for cognition-driven senti- ment analysis,” in Proc. 28th AAAI Conf., 2014, pp. 1515–1521.

[28] E. Cambria, S. Poria, and R. Bajpai, “SenticNet 4: A semantic resource for sentiment analysis based on conceptual primitives,” in Proc. Int. Conf. Comput. Linguistics, 2016, pp. 2666–2677.

[29] K. H. Lin, C. C. Yang, and H. Chen, “Emotion classification of online news articles from the reader’s perspective,” Web Intell., vol. 1, pp. 220–226, 2008.

[30] K. H.-Y. Lin, C. Yang, and H.-H. Chen, “What emotions do news articles trigger in their readers?” in Proc. 30th Annu. Int. ACM SIGIR Conf. Res. Development Inf. Retrieval, 2007, pp. 733–734.

[31] T. Xu, Q. Peng, and Y. Cheng, “Identifying the semantic orienta- tion of terms using S-HAL for sentiment analysis,” Knowl.-Based Syst., vol. 35, pp. 279–289, 2012.

[32] V. Stoyanov and C. Cardie, “Annotating topics of opinions,” in Proc. 6th Int. Conf. Lang. Resour. Eval., 2008, pp. 3213–3217.

[33] D. M. Blei, B. B. Edu, A. Y. Ng, A. S. Edu, M. I. Jordan, and J. B. Edu, “Latent Dirichlet allocation,” J. Mach. Learn. Res., vol. 3, no. 1, pp. 993–1022, 2003.

[34] M. Rosen-Zvi, T. Griffiths, M. Steyvers, and P. Smyth, “The author-topic model for authors and documents,” in Proc. 20th Conf. Uncertain. Artif. Intell., 2004, pp. 487–494.

[35] D. Ramage, D. Hall, R. Nallapati, and C. D. Manning, “Labeled LDA: A supervised topic model for credit attribution in multi- labeled corpora,” in Proc. Conf. Empirical Methods Natural Language Process., 2009, pp. 248–256.

[36] D. Ramage, S. Dumais, and D. Liebling, “Characterizing micro- blogs with topic models,” in Proc. 4th Int. AAAI Conf. Weblogs Soc. Media, 2010, pp. 130–137.

[37] C. Lin and Y. He, “Joint sentiment/topic model for sentiment ana- lysis,” in Proc. Conf. Inf. Knowl. Manage., 2009, pp. 375–384.

[38] A. Balahur, J. M. Hermida, and A. Montoyo, “Building and exploiting E MOTI N ET , a knowledge base for emotion detection based on the appraisal theory model,” IEEE Trans. Affect. Comput., vol. 3, no. 1, pp. 88–101, Jan.-Mar. 2011.

[39] C. O. Alm, “The role of affect in the computational modeling of natural language,” Lang. Linguist. Compass, vol. 6, no. 7, pp. 416– 430, 2012.

[40] S. Poria, E. Cambria, G. Winterstein, and G. Bin Huang, “Sentic patterns: Dependency-based rules for concept-level sentiment analysis,” Knowl.-Based Syst., vol. 69, no. 1, pp. 45–63, 2014.

[41] S. Poria, I. Chaturvedi, E. Cambria, and A. Hussain, “Convolutional MKL based multimodal emotion recognition and sentiment analysis,” in Proc. IEEE 16th Int. Conf. Data Mining, 2016, pp. 439–448.

[42] N. Majumder and I. P. Nacional, “Deep learning based document modeling for personality detection from text,” IEEE Intell. Syst., vol. 32, no. 2, pp. 74–79, 2017.

[43] T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estima- tion of word representations in vector space,” Proc. Int’l Conf. Learning Representation, pp. 1–12, 2013, arXiv:1301.3781.

LI ET AL.: PREDICTING SOCIAL EMOTIONS FROM READERS’ PERSPECTIVE 263

Authorized licensed use limited to: University of the Cumberlands. Downloaded on September 25,2021 at 02:23:41 UTC from IEEE Xplore. Restrictions apply.

[44] S. S.-M. S. Kim, E. Hovy, S. S.-M. S. Kim, E. Hovy, and E. Hovy, “Determining the sentiment of opinions,” in Proc. 20th Int. Conf., 2004, Art. no. 1367–es.

[45] R. E. D. L. Lopes and O. Vian, “The language of evaluation: Appraisal in English,” in Proc. Symp./Workshop Electron. Des. Test Appl., 2007, pp. 371–381.

[46] A. Moors, P. C. Ellsworth, K. R. Scherer, and N. H. Frijda, “Appraisal theories of emotion: State of the art and future devel- opment,” Emot. Rev., vol. 5, no. 2, pp. 119–124, 2013.

[47] B. Meuleman and K. R. Scherer, “Nonlinear appraisal modeling: An application of machine learning to the study of emotion production,” IEEE Trans. Affect. Comput., vol. 4, no. 4, pp. 398–411, Oct.-Dec. 2013.

[48] J. W. Pennebaker, M. R. Mehl, and K. G. Niederhoffer, “Psychological aspects of natural language. use: Our words, our selves,” Annu. Rev. Psychol., vol. 54, no. 1, pp. 547–577, 2003.

[49] B. Smith, “Gestalt theory: An essay in philosophy,” in Proc. Found. Gestalt Theory, 1988, pp. 11–81.

[50] A. M. Czopp, A. C. Kay, and S. Cheryan, “Positive stereotypes are pervasive and powerful,” Perspect. Psychol. Sci., vol. 10, no. 4, pp. 451–463, 2015.

[51] F. Arendt, “Dose-dependent media priming effects of stereotypic newspaper articles on implicit and explicit stereotypes,” J. Com- mun., vol. 63, no. 5, pp. 830–851, 2013.

[52] W. T. L. Cox and P. G. Devine, “Stereotypes possess heteroge- neous directionality: A theoretical and empirical exploration of stereotype structure and content,” PLoS One, vol. 10, no. 3, 2015, Art. no. e0122292.

[53] T. Mikolov, I. Sutskever, K. Chen, G. Corrado, and J. Dean, “Distributed representations of words and phrases and their compositionality,” in Proc. 26th Int. Conf. Neural Inf. Process. Syst., Oct. 2013, pp. 3111–3119.

[54] M. Jacomy, T. Venturini, S. Heymann, and M. Bastian, “ForceAtlas2, a continuous graph layout algorithm for handy net- work visualization designed for the Gephi software,” PLoS One, vol. 9, no. 6, 2014, Art. no. e98679.

[55] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, “Fast unfolding of communities in large networks,” J. Stat. Mech. Theory Exp., vol. 0008, no. 10, pp. 155–168, 2008.

[56] G. E. Hinton and S. Roweis, “Stochastic neighbor embedding,” in Proc. Int. Conf. Neural Inf. Process. Syst., 2002, pp. 857–864.

Xintong Li is currently working toward the MS degree in Xi’an Jiaotong University. His research interests are affective computing, data mining, bioinformatics and affective text mining.

Qinke Peng received the BS degree in applied mathematics, the MEng and PhD degrees in sys- tem engineering from Xi’an Jiaotong University, China, in 1983, 1986 and 1990, respectively, and visited the University of Bordeaux, France in 1994 and University of Oxford, England in 2000 as a research scholar. Currently, he is dean of the Department of Automation Science and Tech- nology as well as a professor of System Engi- neering Institute, Xi’an Jiaotong University. His research interests include modeling and optimi-

zation of system engineering, platform development for big data mining and machine learning, bioinformatics, optimization and security analysis for cyber-physical systems.

Zhi Sun is currently working at Xi’an Jiaotong University for the PhD degree. His main research area is data mining and machine learning.

Ling Chai is currently working at Xi’an Jiaotong University for the MS degree. Her main research area is Data mining.

Ying Wang received the BS degree in electrical and computer engineering by the accelerated program, in 2010, and the MS degree in system engineering from Xi’an Jiaotong University, in 2013. She is currently employed at Xi’an Jiaotong University as a faculty member. Her research interests include the area of data mining and algorithm methods.

" For more information on this or any other computing topic, please visit our Digital Library at www.computer.org/publications/dlib.

264 IEEE TRANSACTIONS ON AFFECTIVE COMPUTING, VOL. 10, NO. 2, APRIL-JUNE 2019

Authorized licensed use limited to: University of the Cumberlands. Downloaded on September 25,2021 at 02:23:41 UTC from IEEE Xplore. Restrictions apply.