The PREventive Care Infrastructure based On Ubiquitous Sensing (PRECIOUS) project aims to develop a preventive care system to promote healthy lifestyles. One of the goals of the project is the development of a method and application for automatic identi cation of human mood. To this end, we hypothesize that, in addition to using smart pervasive artifacts, leveraging in uential factors from social media signals for inferring individuals' moods may enhance the performance of the mood prediction process and furthermore, may reduce the total sparsity and uncertainty of information regarding this process. Accordingly, this position paper describes how our experiment was conducted and reports on our primary achievements for the development of a mood predictor of social media data. Furthermore, we report on the development of a wearable app which enables us to collect explicit feedback from users for conducting a study to evaluate our approach.
The PREventive Care Infrastructure based On Ubiquitous
Sensing (PRECIOUS) project 1 aims to develop a
preventive care system to promote healthy lifestyles. It fconsists of
three components: (
One of the goals of the project is the development of a method and application for automatic identi cation of human mood, as the quality of emotions contributes significantly to the eating habits [5]. Pursuing this goal, sensors and applications of new smart pervasive artifacts (such as smart-phones and smart-watches) can capture diverse spatio-temporal data about an individual from various sensors and applications. The resulting personal life stream data can support powerful inference with regard to the individual's moods and behavior [8].
However, the main challenges of running personal life stream data collections and context-sensing applications are high energy consumption, uncertainty and sparsity of information. Many of these applications that require context information may occasionally need continuous or frequent context monitoring. On the other hand many users use social media platforms for social interactions and express their moods and activities via textual communication and social interactions. These also provide useful signals about the individual. To this end, we hypothesize that leveraging inuential factors from social media communications for inferring moods may reduce the total sparsity of information and uncertainty of mood identi cation process.
For this purpose, this position paper describes the manner in which our experiment was set up and our primary achievements for the development of a mood classi er of social media data. We developed classi ers for moods, \Happy" and \Sad". In addition, as one of the main requirements of the project, we also developed a classi er for \Stress". We predict a ective states from explicit mood-oriented sentences, collected via a mechanical turk study. These a ect-labeled posts are used in a classi cation experiment to predict the a ective state from posts. Our experimental results indicate a wide variation in classi er performance across di erent affects and classi cation algorithms { this may be the result from how patterns and styles when using language vary depending on a ective states.
Recently, some works have investigated development of mood prediction models and methods from individuals' social media communications. For instance in [3] a web-based tool `MoonPhrases' was created to enable Twitter users to reect about their mood and well-being. Therefore the Tweets of a user are analyzed with LIWC (Linguistic Inquiry and Word Count2) to get information about positive-, negative a ect and linguistic styles. These results were visualized with moons and plain text to enable the user to re ect about his/her mood in the past. A similar approach was taken in [4], it was investigated to improve the classi cation of Tweets in either positive, neutral or negative sentiment. Therefore three classi ers were used, then not the results but the probabilities (or con dences) of the classi ers for each class were compared. Finally the class with the highest average probability is chosen.
Moreover in [6] and [2] messages of Twitter users were interpreted to nd out how it is talked about depression in Tweets and how the usage of sentiment words of a depressed person di er from a not depressed person. They have shown that depressed persons show lowered social activity, they post more about themselves and interact less with others. Also depressed persons seem to express more negative emotion with the use of words of certain a ect categories such as anger, causation, tentative, communication, and friends. Furthermore Park et al. [6] revealed that depressed person tweet much personal information about their depression. This is also in agreement with the ndings of Choudhury et al. [2] where they noticed a higher incidence of Tweets dealing about medical concerns. Moreover depressed users tweet more about relational concerns and religious thoughts than not depressed persons and tend to have smaller, tightly clustered close-knit networks with other users on Twitter. Furthermore in [11] rst a qualitative analysis of past Tweets about health or tness was done. As a result a taxonomy was created, which classi es the post according to its activity and the sentiments expressed. Second qualitative interviews were done with Twitter users to nd out about the motivation of posting. The interviews have shown that all participants didn't actively search for health or tness related content. They came across the tness community in Twitter and explored the content of the Tweets but didn't actively post in the beginning.
Nevertheless, our main focus in this project is leveraging social media communications for solving uncertainly and sparsity issues of data for prediction of individuals' mood.
In order to investigate our hypothesis, we have developed an evaluation framework. The framework:
Collects activities of individuals via smart pervasive
artifacts using the open source applications UbiqLog [10]
implicitly. The framework collects individual users'
activities and social interactions in periodical rhythms:
(
Compares and evaluates performance of mood predictions in three di erent settings using explicit feedback collected from individuals. 3.1
To create a ground-truth dataset from real world social media text communications, we performed a mechanical turk study3. We asked each turker to provide us ve of their Facebook posts related to "Sad", "Happy", and "Stress" moods. In order to ensure the quality of the work by coders, we requested them to provide the Facebook pro le address of popular persons (singer, sportsperson, etc.) and provide us ve posts of the selected person with \Happy" tone. The goal was to detect possible inconsistencies and ensure that answers were speci c and not given randomly. 100 turkers participated in this study and only those who had already received a total of Human Intelligence Tasks (HITs) higher than 5000 and HIT Approval Rate higher than 98% were accepted. In total, we collected 1500 posts related to three moods (500 happy, 500 sad, and 500 stress). It is important to note that to be allowed to participate turkers had to be active users of social media platforms, such as Facebook, and provide us their user ID.
For developing mood classi ers, we rst set up two sets of features for classi cation: 1. Text-based Features (TB): We use a standard classi cation feature setup that is common in text and sentiment classi cation. Posts are represented as vectors of unigram and bigram features. Before feature extraction, posts are lowercased, URLs are removed, and numbers are normalized (canonical form). Next, feature reduction takes place. First, features that occur fewer than ve times are removed. Second, features are subsequently reduced to the top 120 features in terms of log likelihood ratio. 2. Psycholinguistic Features (PLB): We utilized an established source of text analysis dictionary, Linguistic Inquiry and Word Count (LIWC), to develop a set of features. LIWC was demonstrated by previous work [7] as a useful resource for identifying emotions of user-generated content. For LIWC, we used a ectiveindicative categories like positive/negative emotions, anxiety, sadness, and anger.
Subsequently, we chose three classi er algorithms to evaluate
their performance for mood identi cation: Logistic
Regression (LR), Support Vector Machine (SVM), and Bayesian
network (BN) classi er. Also, for developing the mood
classi er, we used two modeling approaches: (
Classi cation results for di erent modeling approaches and various combinations of features are given in Table 1-3. The results demonstrate the e ectiveness of using text-related features for inferring individual moods. Nevertheless, for all moods, training a classi cation model using both sets of features shows improved performance compared to the same models trained using one set of features. Surprisingly, we observe that the Bayesian network classi er performs best for almost all moods with di erent combinations of features. The best performances are observed for the mood \Stress", while the worst are for \Sad". More precisely, in the case of the \Stress" mood, we are able to achieve an F1 score of 0.88, coupled with high precision and recall, when using the Bayesian network classi er in combination with all the features. Similarly, for the same setting, we achieve an F1 score of 0.86, coupled with also high precision and recall for \Happy" mood. However, we nd a lower level of F1 score (0.79) when using the same classi er for \Sad" mood, but it is still the best performance setting for this mood. With regard to binary or multi class modeling, we observe that the binary class classi er using both sets of features outperforms other models and, in particular, outperforms models with binary classes using only one set of features. As the text-related features play an important role for mood classi cation, we computed a ranked list of terms from a set of 1500 posts for each mood (500 posts for each mood) as an illustrative example. For ranking the terms, we used the Mutual Information (MI) measure from the information theory which can be interpreted as a measure of how much the joint distribution of features Xi (terms in our case) deviate Features TB PLB Both Features TB PLB Both
Classifier LR SVM BN LR SVM BN LR SVM BN from a hypothetical distribution in which features and categories are independent of each other. Table 4 shows the top 20 terms extracted for each category. Obviously, many of the \Happy" posts contain terms expressing sympathy or commendation. \Sad" posts, on the other hand, often contain negative adjectives.
In order to evaluate and compare usage of di erent signals from various channels we require explicit feedback from the user. Therefore, we implemented a simple smartwatch data collection application (Figure 1). Both smartphones and wearable devices (e.g. smartwatches) can be used to collect information on human behavior. However, wearables have a higher potential for gathering more personal information due to their close proximity to users. In particular, devices such as tness trackers and smartwatches have two major advantages over smartphones, being constantly connected to the skin and located on the body. As a result, these features make them more capable than smartphones of collecting physiological and explicit data [9].
The explicit feedback collector app collects users' moods and locations as explicit inputs within the users' physical activities as implicit inputs. In a more technical sense, if the user shakes the watch, then a pop-up appears and allows them to enter their mood manually. This app enables users to Features TB PLB Both
Classifier LR SVM BN LR SVM BN LR SVM BN explicitly enter their current mood, location (home, work, leisure), and activity four times per day. Physical activity terms are inspired by Google Play services and mood annotation terms are derived from the Circumplex a ect model, which contains two orthogonal dimensions: pleasure (from sad to happy), and activeness (from sleepy to aroused).
In order to investigate the reduction of the total sparsity
of information and uncertainty of the mood identi cation
process exploiting social media text communications, this
paper describes our experimental set up and our primary
achievements for the development of a mood predictor.
Furthermore, it reports the development of a wearable app for
collecting explicit feedback from users in order to evaluate
our approach. We will work on the following steps in future
work: (
The PRECIOUS project has received funding from the European Community's Seventh Framework Programme for research, technological development and demonstration under grant agreement no 611366.
Figure 1: User interface of the prototype implementation of explicit feedback collectors on smartwatch