Measuring Controversy on Social Media
https://dl.acm.org/doi/abs/10.1145/3140565
The paper “Quantifying Controversy on Social Media” discusses how the most controversial topics were detected on social media using the social media network structure and its content. This controversy detection was generalized to identify topics in all domains rather than specific topics in certain domains. Their approach to controversy detection involves a three step process where they first build a conversation graph on the topic, then partition the graph, and finally measure controversy based on the partitioned graph’s characteristics in each partition. They measure controversy through random walks on graphs based on the idea that on each end of the controversy graph, there are opposing opinions, with any nodes between these two ends being users exposed to the topic of controversy.
The Random Walk Controversy (RWC) measure was defined as the difference in probabilities of two events, the first being the probability that both random walks started in the partition that they ended in, and the second being the probability that both random walks started in a partition that is different from the one they ended in. Essentially, the measure is a difference of two values that are each a product of conditional probabilities.
(RWC = P_XX * P_YY – P_YX * P_XY) where P_AB is the conditional probability that the random walk starts in partition A given that it ends in partition B.
This RWC measure was used to determine the controversy of an entire conversation graph, so another measure that determines the controversy of an individual within the graph is an adaptation of the RWC measure above. This RWC measure, denoted by RWC^user, uses Bayes’ theorem.
Given users u and vertices X and Y, with X^+ and Y^+ being high-degree vertices, the RWC measure of an individual is defined by
RWC^user (u, X) = Pr[start = u | end = X^+] / ( Pr[start = u | end = X^+] + Pr[start = u | end = Y^+] )
The content discussed in this section of the paper utilizes both graph theory and probability to detect controversy. In the end, this system was able to identify these controversial hashtags in a three month period between June 25, 2015 and September 19, 2015 on Twitter: #whoisburningblackchurches with a score of 0.332, #communityshield with a score of 0.314, and #nationalfriedchickenday with a score of 0.393. All of these hashtags were trending topics with two opposing sides. It was also determined that the random walk based measure was able to effectively separate controversial from non-controversial topics through the partitioning of the conversation graphs.