Most of our online signals on social media can reveal our views!
How do online social interactions reveal unexpressed views?
This post summarizes the CSCW 2019 paper “ Your Stance is Exposed! Analysing Possible Factors forStance Detection on Social Media” by Abeer ALDayel and Walid Magdy.
The purpose of our research is to understand the online signals that can reveal the stance/views of social media users. This paper analyzes various online signals of users to detect their stance towards politics, religion and other topics. We compare multiple sets of online signals, including on-topic content, network interactions, user’s preferences, and online network connections. Our findings show that unconscious online interactions are enough to predict the unexpressed views on social media.
People use social media platforms to express their viewpoints towards various topics and events. This kind of dependency makes social media a valuable source to harvest social data to discover what are the current public opinion towards an event. The studies in this realm tend to use various online signals ranging from the textual content to the online network connections of these platform users. Yet the question remains, do social media users have to express their viewpoints explicitly to be predictable? What can the unconscious online social signals of users’ reveal about their attitude towards a topic?
In our study, we carried out an analysis of different online signals to evaluate the predictability of their stance towards multiple topics. We categorized the online signals based on two common user’s online behaviours as follows:
1- Active users: the active users on social media tend to post and interact directly by using mentions, reply or retweet. For this kind of users, we used the “Activity Network” and collected the tweets from their home-timeline.
2- Silent users: these users tend to like and follow other accounts without explicitly posting any content. For those, we used the “Preference Network” to collect the mentions and URLs from the tweets they like; in addition to their “Connection Network” which consists of the set of accounts they follow.
We built a prediction model for stance and used different groups of online social signals as features. Our models were tested on the popular stance detection SemEval dataset that consists of over 4000 tweets covering five different topics. For each of the tweets we collected the timelines of the user posted them and the list of tweets they liked. We examined the performance of stance prediction using different sets of features, including the user’s activity, preference, and connection networks and compared it to the state-of-the-art results that use the tweets text for prediction.
The findings show that the online signals left by social media users have enough power to predict the stance. The overall performance produced by each of the networks of users is better than using the textual cues derived from the user explicitly expressing their views. Using the preference or connection networks helps in revealing the silent user who do not post anything online. Additionally, when we combine the network signals along with the content, the stance detection models achieved the best reported performance on the SemEval dataset with an F-Score of 72.5%.
Which online signals reveal the stance?
To understand the influenceability of each online signals on the polarized stance, we derived the most influential factors from the stance detection model. For each of the five topics, we show that most of the online interactions can be irrelevant to the topic itself, as shown in figure 1 and 2. For instance, users who support the feminist movement tend to interact with @MTV. Interestingly, the context of the interaction is also off the main topic of the stance (figure 3). Additionally, interaction with opponent users helps in predicting the views of the users. For example, people with a stance supporting Hillary Clinton tend to interact with Donald Trump account. This kind of interaction is not for supporting Trump, as shown in figure 3. Overall, the interactions with news outlets are the most influenceable factor in predicting the stance. The @telegraph is one of the top factors in predicting the support stance to climate change. On the other hand, people with stance against climate change tend to interact with @SkyNewsBreak .
The findings of our study show the vulnerability of users on social media platforms where their views can be easily predicted. Our work demonstrates how users’ stances can easily be detected even without having the user’s explicitly discussing the topic or even posting at all online. To this end, it becomes highly essential to develop methods and regulations to help in preserving the privacy of the users of social media.
If you want more details on this study, consider further reading & citing below paper. For any questions, please feel free to reach out to the authors.
Abeer AlDayel and Walid Magdy. Your Stance is Exposed! Analysing Possible Factors forStance Detection on Social Media. 2019. CSCW. ACM, Austin, TX. 20 pages.
