Paper 1: Bond, Robert M., et al. “A 61-million-person experiment in social influence and political mobilization.” Nature 489.7415 (2012): 295.

Paper 2: Kramer, Adam DI, Jamie E. Guillory, and Jeffrey T. Hancock. “Experimental evidence of massive-scale emotional contagion through social networks.” Proceedings of the National Academy of Sciences (2014): 201320040.

Bond and Kramers papers on social influence analyze how sentiment and movement is inspired though the social network. Bond et. al. study how much effect a peer network has on the voting action and interests in over 60 million people. Kramer et. al. study how positive and negative posting behavior stems from from seeing other users positive or negative expressions on social media. While both papers study how influence is shaped on social media, they present contrasting results.

Bond et. al. say that peer influence is more important that impersonal nudging. For example, people who saw their friends voting, were more likely to open the link to the nearest polling station. Also they were more likely to engage in self expression and offline action. Even though the data is massive, there is no measure of approximately how many people who select “I voted” sticker, actually vote. Also, they mention that clicking on the link pointing to the nearest polling station was considered as expressing additional interest. Most of the times, the polling stations are locally very well known and frequented by the voters. For example, schools, supermarkets, community places etc. Therefore, it is possible that users who did not click on it, simply did so because they already had the information they needed. To avoid and even identify such limitations, qualitative study of representative sample of users could have been done. This study could have included surveying for reasons of voting and whether their Facebook friends had anything to do with it. Further, this whole paper studies social influence through only one issue: voting. Voting in general is considered as a relatively low trouble and positive activity. It will be interesting to see what effect peer influence can have on activities that take some effort such as vaccination, donation, cleaning drives and awareness campaigns.

The second paper makes seemingly simar but still contrasting claims to the first paper. Kramer el. al. claim that emotions are transferred over social media just by being exposed to others emotions rather than the actual event. This might suggest that a few strongly opinionated users might be deciding the emotional state of their entire peer network. Social media can be said to be lead by virtual extroverts. This makes it even more important for users to get exposed to cross cutting views to avoid getting influenced by a few loud friends. What is different in this paper is their claim about this emotion transfer not being a conscious effort. Firstly, without a good qualitative analysis, strong causality between emotion transfer can’t be made. Secondly, in Bond et. al.’s work, suggests that users are motivated by their friends decision to vote this indicating their conscious decision to value their friends actions over strangers.

Both of the experiments were interesting and had massive dats. I would be interested in discussing the ethical side of their data collection effort and what is reflects about privacy on social media.

Read More

Paper 1: Danescu-Niculescu-Mizil, Cristian, et al. “A computational approach to politeness with application to social factors.” arXiv preprint arXiv:1306.6078 (2013).

This papers analyzes Wikipedia and StackExchange requests data for politeness markers based on politeness theory and builds a predictive model to judge politeness index of a conversation. There are several strong points about this paper such as data filtering, balanced analysis, and feature justification. I specifically liked the part where strong and weak politeness annotations based on percentile agreements.  I also like the application of politeness to power and status.

Even though the paper claims to be domain independent, what is considered to be polite can change across communities. Very obvious and extreme examples of this can be 4chan and Wikipedia. StackExchange and Wikipedia are both information-driven communities in which conversations are situated on a much different level than other social networks such are Twitter and Facebook. Further, politeness can also change based on the context of the social groups it is analyzed in. For example, what is considered as polite in a group of strangers might differ from politeness norms in a group of friends or colleagues.

Few questions come up after reading the annotation procedure of the paper.

1. Is it realistic to get such perfect 50-50 distribution of politeness scores in both datasets? How both datasets have almost an equal number of polite and impolite requests?  (It might be a coincidence, just looks a bit weird!)
2. The median and variance can depend on how many sample points are chosen. It’s not clear how many randomly scored annotations were used for comparison in Wilcoxon signed rank test.
3. Further, the median of both randomized scores is the same. (Again, might be the coincidence but looks weird! )

Further, the author’s findings that failed candidates get more polite after elections are inconsistent with several works analyzing the effect of the community feedback on user behavior. Considering failure in the election as a negative feedback, users should be expected to get more hostile/deviant. This raises a question: Does the correlation of politeness with power change over incentive driven and punishment driven communities?

Using conversation analysis principles in politeness measure: Even though the turker study was conducted with the context of request and answer pair, the linguistic measures were performed only on the request texts. Politeness is conversation driven. It can change over a series of message exchanges based on how the conversation flows. Principles of conversation analysis can be applied here to get a deeper understanding of politeness in conversation rather than just a part of a text. Specifically, talk-in-turns, attempts to repair problems and action formation aspects of conversation can be important in identifying politeness.

Lastly, the work done in this paper can be used to construct the model for assessing the general concept of “status” on social media. This work proves that politeness and status are correlated. Other aspects of status such as respect, authority, credibility, decisiveness can be assimilated to form a social status model. Such a model, as described in earlier class sessions can nudge people to be more source aware, civil, polite and respectful based on the status index in that community.

 

Read More

Paper 1: Starbird, Kate. “Examining the Alternative Media Ecosystem Through the Production of Alternative Narratives of Mass Shooting Events on Twitter.” ICWSM. 2017.

Startbird et. al’s paper qualitatively analyzes sources of news related to shooting events on twitter based on the interpretive graph analysis. She analyzes how alternative sources and mainstream sources are referred in different contexts when it comes to conspiracy theories. That is, either in support or to dispute the claims. The paper also dwells on how such citing of alternative news sources can foil politically motivated conspiratorial thinking.

One common theme that is present throughout the paper, is how all of the events are blamed on one central “untouchable” entity such as U.S government or a secret organization. This comes from a very common conspiratorial trait according to which “everything needs to have a reason”. It is found that conspiracists try to relive the event as an aftermath of that event by “rationalizing” it.[1] Further, such theories go on to giving increased importance to certain agents such as FBI, government, Illuminati etc. The paper mentions that 44 out of 80 sources were promoting political agenda. It would be interesting to know which agents such source frequently target and how do they tie such agents to multiple events. 

The paper makes another excellent point that “exposure to online media correlates with distrust of mainstream media”.  Considering that the mainstream media is correcting conspiratorial claims or presenting the neutral narrative, it will be interesting to do a contrast study in which network of users “bashing” mainstream media can be found. One important thing to note here is that just by text analysis methods, it is difficult to understand the context in which the source of information is cited. This reflects in the three ways mentioned in the paper by which alternative media promotes alternative narratives. 1. They cite alternative sources to support the alternate narrative or 2. They cite mainstream sources in a confrontational way. This is where the quantitative approaches are tricky to use because clearly, just the presence of a link in a post doesn’t tell much about the argument made about it. Similarly, hand-coding techniques are also limiting because analyzing the context, source and narrative takes a long time and results in high quality but smaller dataset. One possible way to automate this process can be to perform “Entity Sentiment Analysis” that combines both entity analysis and sentiment analysis and attempts to determine the sentiment (positive or negative) expressed about entities within the text. Treating the sources cited as “proxy” entities, it can be possible to find out what is being said about them in the positive or negative light.

The paper also mentions that believing in one conspiracy theory makes a person more likely to believe another. This, alongwith the cluster of sources supporting alternative narratives domains in the figure 2 can form a basis for quantitatively analyzing how communities unify. [2]

Lastly, as a further research point, it is also interesting to analyze when a particular alternate narrative gets popular. Why some theories take hold while many more do not? Is it because of the informational pressure or because of the particular event. One starting point for this kind of analysis is [3] which says that when the external event threatens to influence users directly, they explore content outside their filter bubble. This will require retrospective analysis of posting behavior before and after a specific event considering users which are either in geographical, racial any other ideological proximity of the group affected by that event.

 

 

[1] Sunstein, C. R., & Vermeule, A. (2009). Conspiracy Theories: Causes and Cures. Journal of Political Philosophy, 17(2), 202–227. https://doi.org/10.1111/j.1467-9760.2008.00325.x

[2]Science vs Conspiracy: Collective Narratives in the Age of Misinformation. PLOS ONE, 10(2), e0118093. https://doi.org/10.1371/journal.pone.0118093

[3] Koutra, D., Bennett, P., & Horvitz, E. (2015). Events and Controversies: Influences of a Shocking News Event on Information Seeking. TAIA Workshop in SIGIR, 0–3. https://doi.org/10.1145/2736277.2741099

Read More

Video Reflection #9 – [09/27] – [Shruti Phadke]

The effect of cognitive biases on information selection and processing is a well-established phenomenon. According to Eli Pariser, the person who coined the term “filter bubble”:

“A world constructed from the familiar is a world in which there’s nothing to learn … (since there is) invisible auto propaganda, indoctrinating us with our own ideas.”

Some previous research has been put into how to nudge the reader towards more cross-cutting content. For example, ConsiderIt exposes users to the choices made by others to provoke perspective re-evaluation. Similarly, OpinionSpace encourages users to seek out diversity through the graphical reflection of their own content. Other than this, the formation of diverse views mainly depends on “serendipitous discovery”, the conscience of mainstream media or the user’s willingness to accept the opposing content. Dr. Stroud mentions that selectivity can be influenced using forced exposure and prompting for accuracy or informing users of their filter bubble. But, is just “exposure” to anti attitudinal content enough to promote real diversity? Does is matter how the information is framed? Jonathan Haidt’s Moral foundation theory interests me the most in this regard. Haidt and colleagues found that liberals are more sensitive to care and fairness while the conservatives emphasize more on the loyalty, authority, and sanctity.  This raises a question of whether a conservative reader can be encouraged to read a left-leaning news by using words associated with conservative moral foundations? Similarly, will a liberal entertain conservative thoughts simply if they highlight fairness and empathy? Stroud’s findings in the second research present strengthen the argument. She reports observing that newsrooms encourage controversy. Also, partisan comments attract more incivility. This might make newsrooms a discouraging place to get exposed to cross-cutting content especially if it is associated with unfavorable framing.

This can form a basis for an experiment that studies how the framing of information affects the acceptance of a cross-cutting content. This research can be done in the collaboration of linguistic experts who can attest to various framings of a similar information consistent with the moral foundations of the user group. Participants can be self-identified liberals and conservatives not getting exposed to the differently polarized news. A control group can consist of users who are exposed to cross-cutting content without reframing the news/information. There can be two treatment groups with the following treatments
1. Exposure to cross-cutting content with conservative moral framing
2. Exposure to cross-cutting content with liberal moral framing

Finally, the effect can be observed in terms are how likely conservatives/liberals are to select a cross-cutting information which is wrapped up in a language corresponding to a specific moral foundation. Further, instead of limiting the study to conservative/liberal moral foundations, the experiment can also explore the effect of all moral foundation dimensions. (care, fairness, loyalty, authority, sanctity)
This type of study can inform what makes cross-cutting news more appealing to specific users and how it can promote diverse ideologies.

 

Read More

Paper 1: Erickson, Thomas, and Wendy A. Kellogg. “Social translucence: an approach to designing systems that support social processes.” ACM transactions on computer-human interaction (TOCHI) 7.1 (2000): 59-83.

Paper 2: Donath, Judith, and Fernanda B. Viégas. “The chat circles series: explorations in designing abstract graphical communication interfaces.” Proceedings of the 4th conference on Designing interactive systems: processes, practices, methods, and techniques. ACM, 2002.

Social cues and signals are the most important part of any communication. Erickson et. al.’s paper discusses how subtle details and peculiarities of offline communication can improve the coherence of the online communication. They categorize the present digital conversational systems in three parts–Realist, Memetic and Abstract. Donath et. al.’s paper presents one such abstract concept–The chat circle series–to illustrate how graphical interface can make online systems more engaging and sociable. Since both the papers are considerably old, I am structuring my review as the comparison between points they make and the current systems.

1. Realist Platforms: The realist systems consist of combination of physical and digital signal, for example, video conferencing, teleconferencing etc. Authors mention that such systems are expensive and require infrastructure. But, currently robust and dependable systems such as Zoom, Google Meet, Skype, Facebook Messenger, WhatsApp Video Call  are available for free and can be used on any platform for individual or group conferencing. Some of them also include presentation, side-chatting and commenting tools. Features such as sharing emoticons (hand raising, request for increasing volume, thumbs-up) mimic social cues. Additionally, cross domain realist systems such as watching movies together remotely (https://www.rabb.it) , listening to music remotely (https://letsgaze.com/#/) provide real time multimedia sharing experience. One concern that still remains is that, due to the quality of such video/audio calls, often, subtle expressions and poses go unnoticed.
2. Mimetic Platforms: These platforms include online personas and virtual reality systems where the user has to setup their own avatars to conduct the conversations. Authors mention that it takes a conscious and continuous effort on user’s part to manipulate such systems. With the advance in sersor fusions and Augmented reality, the mimetic systems have traveled a long distance. For example, systems like Room2Room (https://www.microsoft.com/en-us/research/publication/room2room-enabling-life-size-telepresence-in-a-projected-augmented-reality-environment/) are able to facilitate a life sized telepresence for conversations.
    
   Such systems can be very impactful in establishing realistic social cues and interactions, digitally.
3. Abstract Systems: Perhaps the most unexplored area in interactive system is the abstract systems. Donath et. al.’s paper describes “The Chat Circle” series which is designed to increase the expressiveness of the digital communication. The key element of such designs is to enable users to form impressions of the other users based on additional features provided in the graphical interface. Although the design is innovative and take insights from the real world, such designs are not widely used, at-least not yet.

 

[Note: I wanted to write more about using the Chat Circle, but the link is expired now and the system has no access]  

Read More

Paper 1: Sandvig, Christian, et al. “Auditing algorithms: Research methods for detecting discrimination on internet platforms.” Data and discrimination: converting critical concerns into productive inquiry (2014): 1-23.

Paper 2: Hannak, Aniko, et al. “Measuring personalization of web search.” Proceedings of the 22nd international conference on World Wide Web. ACM, 2013.

It is safe to say that the issue of algorithmic bias and personalization was once celebrated as a “smartness” and high service quality of internet information providers. In a study by Lee et. al. [1] the research participants attributed algorithms’ fairness and trustworthiness to its perceived efficiency and objectivity. This means reputed and widely acknowledged algorithms such as Google search appear to be more trustworthy. This makes it particularly severe when algorithms give out discriminatory information or make decisions on user’s behalf. Christian et. al.’s paper reviews the implications of discrimination on internet platforms along with the auditory methods that researchers can use to detect its prevalence and effect. Amongst the several methods proposed, Hannak et. al. in paper two utilize sockpuppet audit and crowdsourcing audit methods to measure personalization on Google search. From the review of two papers, it can be said that the bias in algorithms can occur either from the code or from the data.

1. Bias in code can be attributed to financial motives. The examples given in the paper 1 about American Airlines, Google Health and product ranking highlight this fact. But, how is it different from getting store brand at a low price in any of the supermarkets? On surface, both enterprises are utilizing the platform they created to best sell their products. However, the findings in the paper 2 prove that website ranking (either having AA flight ads at the top or having Google service links at the top) is what separates fair algorithms from the unfair algorithm. (Unlike biased information providers Kroger displays its store brand at the same front as other brands). There is a clear difference between change in search rank between AMT results and the control results.
2. Bias in data, I believe, is mainly caused due to the user’s personal history and the dominance of a particular type of information available. Getting similar type of information based on history can lead to echo chambers of ideologies as seen in the previous paper. There is also another type of bias in data that informs algorithms in the form of word embeddings in automatic text processing.  For example, in the paper “Man is to Computer Programmer as Woman is to Homemaker”, [2] Bolukbasi et. al. state that the historical evidence of embeddings of computer science being closer to male names than female names will make search engines rank male computer scientist web pages higher than female scientists. This type of discrimination can not be blamed on a single entity but just the prevalence of biased corpus and the entire human history!

Further, I would like to comment briefly on the other two research design methods suggested in paper 1. Scraping audit can have unwanted consequences. What happens to the data that is scraped and later blocked (or moderated) by the service provider? Recently, Twitter suspended  Alex Jone’s profile but his devoted followers were able to rebuild a fake profile with real tweets based on the data collected from the web crawlers and scrapers. Also, noninvasive user audits, even though completely legal can be ineffective with poor choice of experts.

Finally, given the recent events, it can be valuable to research how algorithms share information across platforms. It is common to see ads of hotels and restaurants on Facebook after booking flight tickets with Google flights. Is “Google personalization” only limited to Google?  

[1] Lee, Min Kyung. “Understanding perception of algorithmic decisions: Fairness, trust, and emotion in response to algorithmic management.” Big Data & Society 5.1 (2018): 2053951718756684.

[2] Bolukbasi, Tolga, et al. “Man is to computer programmer as woman is to homemaker? debiasing word embeddings.” Advances in Neural Information Processing Systems. 2016.

Read More

Paper 1: Bakshy, Eytan, Solomon Messing, and Lada A. Adamic. “Exposure to ideologically diverse news and opinion on Facebook.” Science 348.6239 (2015): 1130-1132.

Paper 2: Eslami, Motahhare, et al. “I always assumed that I wasn’t really that close to [her]: Reasoning about Invisible Algorithms in News Feeds.” Proceedings of the 33rd annual ACM conference on human factors in computing systems. ACM, 2015

Algorithmic bias and influence on social networks is a growing research area. Algorithms can play an important role in shifting the tide of online opinions and public policies. Both Bakshy et. al’s and Eslami et. al’s papers discuss the effect of peer and algorithmic influence on the social media users. Seeing ideologically similar feed as well as the feed based on interactions can lead to having extremist views and associations online. The “echo-chambers” of opinions can go viral unchallenged within a network of friends and can range from harmless stereotypes to radical extremism. This exposure bias is not just limited to the posts but also to the comments. In any popular thread, the default setting shows only the comments either made by friends or are popular.

Eslami et. al’s work shows how exposing users to the algorithm can potentially improve the quality of online interaction. Having over 1000 friends on Facebook, I barely see stories or feeds from most of them. While Eslami et. al. do insightful qualitative research on how users perceive the difference between “all stories” and “shown stories” along with their future choices, I believe that the study is limited in the number of users as well as different user behaviors. To observe the universality of this phenomenon, a bigger group of users should be observed with users behaviors varying in the frequency of access, posting behavior, lurking, and users with promotional agenda. Such study can be performed with AMT. Even though it will restrict the open coding options and detailed accounts, this paper can serve as a basis for forming a more constrained and precisely defined questionnaire which can lead to quantitative analysis.

Bakshy et. al.’s work, on the other hand, ties the political polarity in online communities to the choices the user has made. It is interesting to understand the limitations of their data labeling process and the content. For example, they have selected only the users that volunteer their polarization on Facebook. Users who volunteer this information might not represent the average population on Facebook. A better classification of such users could have been done by just text classification on their posts without their proclaimed political affiliation. One more reason to avoid their political status is that many users can have a political label attached to them due to peer pressure or the negative stigma attached to their favored ideology in their “friend” circle.

Finally, even though getting exposed to similar or algorithmically influenced content may be potentially harmful or misleading, it also raises the questions about how much data privacy invasion is allowed to de-bias the feed on your timeline. Consciously building algorithms that show cross-cutting content can end up knowing more about a user than he intends. The question of solving this algorithmic influence should be approached with caution and with better legal policies.  

Read More

Kumar, Srijan, et al. “An army of me: Sockpuppets in online discussion communities.” Proceedings of the 26th International Conference on World Wide Web. International World Wide Web Conferences Steering Committee, 2017.

Summary:

Sockpuppets in reference to this paper, are online discussion accounts belonging to the same user, also referred here as the “puppetmaster”. Kumar et. al’s work in studying sockpuppets in online communities observe the posting behavior, interactions and linguistic characteristics of sock puppets, finally leading to a predictive model. They observe that the sockpuppets tend to comment more, write shorter posts, use more personal pronouns such as “I”, and are more likely to interact with each other. Finally, in the predictive task the authors find that activity, community and post features are most relevant to detecting the sockpuppets with 0.68 AUC. Here are some thoughts on the data collection, method and impact of this work:

 

Why is this research even important? Even though this paper has excellent technical and analytical aspects, I believe that there should have been some more stress on why sockpuppetry is harmful in the first place.

“In 2011, a California company called Ntrepid was awarded a $2.76 million contract from US Central Command for “online persona management” operations[42] to create “fake online personas to influence net conversations and spread US propaganda” in Arabic, Persian, Urdu and Pashto.” (Wikipedia)

I found some more reasons which I think are important to situate this research with community betterment

1. Bypassing the ban on the account by creating another account (mentioned in the paper)
2. Sockpuppeting during an online poll to submit multiple votes in favor of the puppeteer.
3. Endorsing a product by writing multiple good reviews
4. Enforcing public opinion about a policy or candidate by sheer numbers

 

How to build a better ground truth? One obvious point of contention with this paper is the way the data is collected and labeled as sockpuppet accounts. There is no solid validation regarding whether the selected accounts are actually sockpuppets. The authors mention that they had conservative filters while selecting the sockpuppet accounts but it also means that they might have missed significant true positives. So what can be done to build a better ground truth?

1. Building a strong “anti-ground truth”. There are performance comparisons between sockpuppets and ordinary users throughout the paper. If the sampled list of ordinary accounts was vetted more strongly (if they had a stronger anti group) the comparisons would have been more telling. One way to do this is to collect accounts which posted from different IPs or location at the exact same time.
2. Asking the discussion groups for sockpuppets. Even though this seems harder, it can form a very strong ground truth and validation point

Lastly, there are several comparisons between the pairs of sockpuppets and two ordinary users. I am not sure whether the ordinary user’s measure was a normalized aggregate of all pairwise ordinary measures. In any case, instead of comparing the sockpuppet pair activity with generic pairwise activity, it would be better to find out the comparison with two ordinary users with some probability of interaction (eg. same discussion, location, posting time etc.) Also, while comparing between pretenders and non-pretenders, it would be beneficial to have a comparison with ordinary users as a ground truth measure.

In the discussion, the authors claim that not all sockpuppets are malicious. Further research can be focused on finding characteristics of only malicious sockpuppets or online deception “artists”!

Read More

Mitra, Tanushree, Graham P. Wright, and Eric Gilbert. “A parsimonious language model of social media credibility across disparate events.” Proceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing. ACM, 2017.

This paper represents an in-depth and meticulous analysis of different linguistic and social network features that affect the credibility of a post. Mitra et. al. take intuition from linguistic cues such as subjectivity, positive emotions, hedging along with social network specifics such as retweets and replies to build a model that maps such features to the level of credibility. This study with thorough experimentation and validation not only provides strong evidence of the effects of such features but also gives qualitative insights and implications of the research.

The model fit comparison section specifically reflects several social network characteristics. For example, getting better explanatory power after including both the original texts and replies highlights the role of context and conversational nature of the social media interaction. Seeing the low predictive power of non lexicon based features, such as hashtags, caps and question marks, I am curious about whether all such features could be grouped into the “readability index” of the corpus corresponding to each event. It is possible that lower readability can be a good predictor of lower credibility. (Although, it is not clear just by intuition whether higher readability will be a good predictor of higher credibility)

Credibility in non-anonymous networks can have strong ties to how the source is viewed by the reader. Authors discuss that they did not include source credibility in the features but I think that “poster status” can also impact the perceived credibility. For example, I am more likely to believe in the fake news posted by my colleagues rather than a stranger with the same source. Similarly, I am more likely to believe in the information provided by a user with higher karma points than one with the lower karma points. Because the credibility annotations were done by turkers, it is not possible to assess the effect of poster status in the current setup. But, in a retrospective study, it is possible to have additional non-lexicon based features such as user statistics and tie strengths between the poster and the reader.

Such analysis that comprises of strong linguistic and non-linguistic features can be also applied to detecting fake news. Websites such as “Snopes”, “PolitiFact”  have pieces of news and the fact-check review on them tagged by “original content”, “fact rating” and “sources” which can be used either for stand-alone analysis or grouping the twitter event streams as fake or credible.

Finally, I believe that consequences of credibility range from disbelieving in scientific and logical information such as the importance of vaccinations and climate change to believing in conspiracy theories and propaganda.  Fast paced online interactions do not allow the users to analyze every piece of information they get. This makes the linguistic and social influence perspective on credibility more relevant and important in de-biasing the online interaction.

Read More

Paper 1: Cheng, Justin, Cristian Danescu-Niculescu-Mizil, and Jure Leskovec. “Antisocial Behavior in Online Discussion Communities.” Icwsm. 2015.

Cheng et. al. in this impressive and thorough work demonstrate a way of observing and predicting antisocial behavior on online forums. They study banned users from three communities, retrospectively, to identify antisocial features of their online activity. Keeping with the Donath’s paper, we can find the similar definition of antisocial behavior here, where users act provocatively posting aberrant content.  One salient strength of this paper is well-defined problem setup and normalizing measures taken before the analysis to compensate for the peculiarities of the social media. For example, they consider strong signals such as banning and post deletion to pre-filter the bad users and also perform qualitative analysis (human evaluation) to assess the quality of posts. Their observation about bimodal post deletion distribution symbolizes the complexity of overall online interactions. It also suggests that antisocial behavior can be innate as well as adaptive. Based on the paper there are several interesting directions that can be explored.

What are the assessment signals within a group of trolls? The paper mentions several signals of antisocial behavior such as down-voting, reporting, blocking, banning etc. These signals are from the point of view of normal users and community standards. However, it will be interesting to observe whether such users identify each other and show group trolling behavior in communities. The paper mentions that the FBUs are able to garner an increased response by posting provocative content. Does this contention increase if there are multiple trolls involved on the same thread? If yes, how do they identify and support each other?

What about the users that were not banned, but are extremely antisocial on occasions? Consider a scenario in which there are two users, U1 and U2. U2 frequently trolls people of different communities by posting harmless but annoying and argument fuelling content. Based on the observations made in the paper, U2 will most likely get banned because of the community bias and post deletion. Now consider U1 who posts rarely but with extremely offensive content. U1 will most likely have their posts deleted without attracting much attention. Comparing by the number of posts deletions (which is the driving factor of this paper’s predictive model) U2 will have more likelihood of getting banned. But, which one of the two (U1 and U2) is actually more dangerous for the community? The answer is, both! To observe both types of behaviors, there should be multiple granularity levels while analyzing antisocial behavior. Per post, per thread and per user. Analyzing hateful or flaming language in individual posts can have some weight in the predictive model for deciding whether the user is antisocial or not.

Finally, will moderators be biased if they are exposed to the prediction results based on the first five posts of the user? In a scenario like this where banning and blocking infringes on the freedom of speech of the user, knowing the likelihood of a particular user being “good” or “bad” might increase community bias and the amount in which these users get censored. So, features based on more linguistic analysis are definitely relevant in which users the moderators should be warned about.

Lastly, there are few more specific questions that I want to mention:

1. Were the Lo-FBUs slower to get banned compared to the Hi-FBUs. Also, did they have a longer/shorter lifespan compared to NBUs and Hi-FBUs. This analysis might give clues about how to build more accurate predictive models for Lo-FBUs.

2. Why is Krippendorff’s alpha so low? (And how was it accepted?)
3. Can we also quantify the sensitivity of the community and use it as a feature for prediction? (Sort of customizing the banning algorithms as per the individual community’s taste. )
4. How to define antisocial behavior in anonymous communities and inherently offensive communities like 4chan?

Read More