Paper-

Data ex machina: Introduction to big data – Lazer et. al.

Summary-

This article was a survey article, talking about the potential of big data in sociology, and computational social science. It served as a very good survey article, providing many examples and references of interesting research being done, which leverages data to glean insight in social science, and talks about the different interpretations of data, depending on the knowledge we wish to gain. The article began with a statement saying that what is captured from data isn’t what the social scientist wants, which makes how to mine what we want from the dump of interactions a difficult task in itself, which is an important theme in the article.

Reflection-

I liked the use of the phrase “substantive substrate of the data is the collective behavior of humans” used to describe the applicability of data to the social sciences, as it paints a good picture of the task of distilling understanding of human behavior from a lot of interactions.

I also found the two interpretations of social media platforms as either microcosms of all of society, versus a realm in itself, interesting. The second interpretation holds that not only are these platforms incomplete at capturing all human experience, but they also modify human behavior in their own right.   

The tools and results obtained by The Copenhagen Network Study were interesting because they tried to obtain meaningful data about interaction from diverse sources, i.e. mobile phone exchanges and facebook, and found that participants were actually using those two communication media for different tasks, to interact with largely distinct groups of friends, reinforcing the second interpretation of social/communications platforms (from the former paragraph).

Another fascinating insight I got from this article was on how big data can be used to cheaply analyze and interpret politically relevant information on a national level, e.g. the research on predicting inflation rates using goods prices, or the research on estimating the impact of a hurricane.

Making big data small, i.e. identifying subgroups of interest within the dataset, like the leaders of a revolution, people with PTSD/other psychological disorders, etc. can be used to study these phenomena retrospectively in an unobtrusive manner.

I also found the term “big data hubris” used in the article interesting, since it helped me understand that volume of data can be misleading if sampling is not done properly, or if you do not understand the data you have. For example, the spiked trends in usage of the word “fuck” in books in the 1800s, in Google Ngrams was found to be due to a failure of OCR systems to read archaic spellings of the letter “s”. The presence of a large number of fake accounts and bots on certain platforms also makes it important to ensure that the data obtained is genuine.

This article was a thought-provoking and fascinating read. It was a wonderful way to conclude the reflections we did in this course, as it gave a high level, but broad insight into research areas in the field of social computing.

Read More

Papers:

[1] Experimental evidence of massive-scale emotional contagion through social networks – Kramer et. al.

[2] A 61-million-person experiment in social influence and political mobilization – Bond et. al. 

Summary:

[1] talks about emotional contagion in social networks; how emotions expressed by others influence our own emotions. It hints that these effects may even be long term, and are often subconsciously experienced. It tries to quantify whether seeing positive posts on their feed makes a user themselves post more positively, and vice versa. 

[2] also talks about the importance of social influence for effecting behavior change, but it focuses on social contagion in actions rather than emotions. It randomly showed users either an informational message (i.e. statistics about how many people voted, and encouragement to vote), a social message (i.e. profile pictures of friends who voted along with the informational message) or no message, and analyzed the effects of the messages on users, and also analyzed the effects of close friends getting the message on users.

Reflection:

The authors in [1] state that they determined a post to be positive or negative if it contained at least one word which was positive or negative according to LIWC respectively, but this seems to be a fairly weak way of measuring the emotional content of the post, and the data might contain lots of “weakly positive” or “weakly negative” posts.

[1] was also an interesting insight into statistical analysis. It used a weighted linear regression model to test the hypothesis. The authors took care to try and reduce the effects of other factors, by ensuring that the subjects had no significant deviations in their emotional expression during the week before the experiment, and conducted t-tests to check the correlation between variables.

However, since emotional state is a complex thing to quantify, and since the sample size was so large, there is a lot of scope for external unforeseen variables which might muddy the results. Since the sample size is large, small effects can be found to have “significant” p-values. Moreover the effects of social contagion were not very pronounced. The percentage of change in number of words of a particular emotion used by a user after changing the news feeds were less than 0.1%.

As the study in [1] was only conducted over a week, studies can also be conducted to see the effects of emotional contagion for shorter or longer periods of time, or measure the rate of change of emotional word usage over time.

In [2], since a major factor in deciding whether users would vote or become interested in voting were the close friends of the user, the results become hard to isolate from effects beyond social media. This is because user’s have plenty of real world interactions with close friends, and this might be what influenced them to vote, not the messages on social media.

Interestingly, in [2], the users who saw the social message were 0.39% likelier to actually vote, but 2.08% likelier to say they voted. This hints that users are aware of the social stigma that comes with not voting, and often the validation from their peers for having voted is a stronger driving force than the inherent desire to vote. Maybe a system can be designed so that only verified voters can click the I Voted button, and the effects that has on voting patterns can be analyzed.

The probability of encouraging a user to vote through online social contagion is found to be quite small, although the authors do (rightly) state that even the small effects can have large consequences, since voting races are narrowly decided. 

Read More

Paper-

A computational approach to politeness with application to social factors – Danescu-Niculescu-Mizil et. al.

Summary-

The authors created an annotated politeness corpus of requests on Wikipedia and Stack Exchange using Amazon Mechanical Turk, which they used to build a classifier which uses domain independent lexical and syntactic features to achieve near human levels of accuracy on classifying a text as polite or impolite. They confirmed some theoretical ideas about politeness, and found that users who were more polite were win adminship elections on these websites, i.e. an editor/admin rank, and that their politeness levels actually decreased after reaching these levels, and the opposite happened for users who also ran for adminship elections and lost. They also analyzed links between status, gender, interaction type, community type, etc and politeness levels.

Reflection-

The authors designed an experiment with both theoretical and practical importance. The fact that the SVM trained on features derived from socio-linguistic/psychological theories of politeness was better than the bag-of-words model, gives more credibility to these theories. Moreover, the model which they trained also has practical importance since it has applications in auto-moderation of online forums, conversational systems, translation systems, etc.

This paper was an interesting read because it taught me a lot about experimental design for computational social science.What struck me about the approach was how cautious and thorough it was. I learned of a number of interesting statistical corrections which the authors applied to get more accurate results, and to ensure the annotation was being done properly, which may be applicable in my project or in future research. For example, they applied z-score normalization to each Turkers score to correct for personal biases in rating a request’s politeness. They also compared pairwise score correlations for Turkers who reviewed the same request, to ensure that the labelling is not happening randomly. Moreover, the authors even trained a “control” SVM based on bag-of-words unigrams to show that the results were due to the power of the features, and not simply the power of the classifier used. These were all great examples of “science done right”, to help budding researchers like myself understand how to obtain and interpret results.

The paper found that question askers were more significantly polite than answerers. Interestingly, high power users were less polite than low power users even when they themselves were asking questions. However, this difference was not extremely pronounced, but it hints at interesting power dynamics/psychological phenomena in the behavior of high power users, which might be worth computationally exploring.

The authors trained their SVM model on Wikipedia, and used it to classify politeness of requests on Stack Exchange. However, this is not a strong enough indicator of the generalizability of this model, since both Wikipedia and Stack Exchange are broadly similar types of sites, (i.e. both deal with acquisition of information). To further prove the robustness of the model, it should be applied to more diverse and unrelated online communities, like 4chan, gaming forums, reddit, etc. 

The model could also be applied to short exchanges beyond just requests, like comment threads, to further see how generalizable the results are.

The authors also mention that the annotated corpus was only made with requests which had two sentences, with the second sentence being the actual request. The use of this corpus is limited, because of its specificity. A more general politeness corpus could be created taking into consideration exchanges of various lengths.

The model used could also be increased in power, i.e. using a deep neural network instead of an SVM for classification, and results could be observed using the new model.

Read More

Paper-

[1] Examining the Alternative Media Ecosystem through the Production of Alternative Narratives of Mass Shooting Events on Twitter – Kate Starbird

Summary-

This paper uses twitter data to construct a network graph of various mainstream, alternate, government controlled, left leaning and right leaning media sources. It uses the structure of the graph to make certain qualitative observations about the alternative news ecosystem.

Reflection-

Some of my reflections are as follows:

This paper uses a very interesting phrase, “democratization of news production”, i.e. it very broadly deals with the fundamental struggle between power and freedom, and the drawbacks of having too much of either. In this case, an increasing democratization of news has weakened the reliability of the very information we receive.

It would be interesting to see what users who are involved in alternative media narratives follow outside the alternative media content, by analyzing their tweets in general, outside of the context of gun violence and mass shootings.

I found the case of InfoWars interesting – it was only connected to one node in the network graph. What were the reasons for that? Maybe infowars did not release much content about mass shootings, or maybe users who use it do not refer to other sources very often, or maybe it just produces content which no one else is producing, and thus sort of stands alone?

Only 1372 users tweeted about more than one alternative media event over the period studied. Maybe another longer-duration study can be conducted, since conspiracy worthy events happen rarely, and 10 months may not be enough to really find the network structure.

It was very interesting that this paper saw evidence of bots propagating fake news, and this paper also later claims that the sources were largely pro-Russia,  which might give some insight into Russian tampering in the 2016 election. The paper also mentions that the sources were not more left leaning or right leaning, but the thing they had in common was an anti-globalist bias, and they insinuated that all of Western Government is basically controlled by powerful external interests, painting the west in a bad light.

The graph provides a sort of spatial link, but it would be interesting to also have a temporal link between source domains, to see what the originators of the information are, and how information propagates over time in these ecosystems. The paper also alludes to this in the conclusion.

The graph is dominated by aqua nodes, which also hints at selective exposure being prevalent here too, providing further evidence to the topic of discussion of last week’s papers, i.e. users who have a tendency to believe in conspiracies, will interact with alternative media more than they will interact with mainstream/other types of media.

It is very interesting that 66/80 alternative media sites cited mainstream media at some point, while not a single mainstream media site cited an alternative media site. It hints at the psyche of alternative media, painting a sort of “underdog” picture for these sources, where they are fighting against an indifferent “big media” machine, which I feel is quite appealing to people who are prone to believing in conspiracies.

The paper states that belief in conspiracies can be exacerbated because someone thinks they have a varied information diet, but they actually have a very one-sided diet. This reminds me of the concept of the Overton Window, which is the name given to the set of ideas which can be publicly discussed at any given time. Vox had a very interesting video on how the Overton Window, after decades of shifting leftwards, is now beginning to shift to the right. This also has an effect on our information diet, where what we feel might be politically neutral, might actually be politically biased, because the public discourse itself is biased. 

Read More

Video-

Partisanship and the search for engaging news – Natalie Stroud.

Reflection-

I found this video particularly interesting, since just last week, my project proposal submission was relevant to selective exposure in online spaces. The idea I had to tackle this problem, especially on platforms which have automatically recommended content, was to create an option for a sort of Anti-Recommender system, which clusters users into groups based on their likes and dislikes, and then serves up recommendations which users in the completely opposite cluster would prefer. This would serve to make people aware about the motivations, arguments, and sources which the “opposite” side has. It could be used not just in politics, but also for a book platform like Goodreads, or even a music platform to help people be exposed to different types of music.

It would also be interesting to explore in more detail the effects of such a system on users; does it incite empathy, anger, curiosity or indifference? does it actually change opinions, or does it make people think of counterarguments which support their beliefs? (this was dealt with in last week’s papers on selective exposure).

Besides analyzing the partisan influences on how people write and interact with comments, it would also be interesting to further break down the categories from two “sides”, down into their constituents, and examine the differences in how the subcategories of these two categories engage with the comments section. For example, how does the interaction vary in both sides, considering minorities, men, women, young, old, etc.

In my opinion, the two keys to understanding selective exposure, and improving how users engage with other users with opposite beliefs are as follows:

1. Understanding the cases where users are exposed to counterattitudinal information, when and why they actively seek it out, and how they respond to it.
2. Designing systems which encourage users to : (i) be more accepting of different viewpoints, and (ii) to critically examine their own viewpoints.

Both of these are of course, addressed in depth in the video. I find that these two areas have huge scope for interesting research ideas, more data analysis driven for point 1, and more design driven for point 2.

For example, a system could be designed which takes data from extensions like the balancer, (which was referred to in the bursting your filter bubble paper from last week), or any similar browser extensions which categorize the political content a person views, and analyze that data to see if a “red” person ever binges on “blue” content for an extended period of time, or vice versa, and identifying any triggers which may have caused this to happen. Historical data can also be collected to find out how these users “used” the data they collected from this binge of counterattitudinal data. That is, did they use it as ammunition for their next comments supporting their own side? were they convinced by it, and did they slowly start browsing more counterattitudinal data?

Similarly, systems can be designed which transform a webpage to “nice-ify” it. This could be a browser extension, which provides little messages at the top of a web-page, reminding users to be nice, or respectful. It could also detect uncivil language, and display a message asking them to reconsider.This ties into the discussion about the effectiveness of priming users to behave in certain ways.

Systems could also be designed to humanize people on chat forums, by adding some (user decided) facts about them to emphasize their personhood, without revealing their identity. It is a lot harder to insult Sarah, who has a 6 month old kitten named Snowball, and likes roller blading, than it is to insult the user sarah_1996. This would also bridge partisan gaps by emphasizing that the other side consists of humans with identities beyond their political beliefs.

Read More

Papers-

[1] Echo chambers online?: Politically motivated selective exposure among Internet news users – Kelly et. al.

[2] Bursting your (filter) bubble: strategies for promoting diverse exposure – Resnick et. al.

Summaries-

[1] discusses how user’s political leanings affect how they interact with news articles. It collects data from hundreds of users of news sites, and conducts a behavioral tracking experiment to see whether users prefer to interact with content they agree with or content they disagree with. It finds that users are less likely to interact with information they disagree with, but they do not actively avoid it. It constructs five hypotheses, considering whether users look at information which supports or detracts from their own viewpoints, and how long they spend looking at these articles.

[2] is a very short survey type paper, which, after quickly defining the need to design tools to provide diverse exposure and discourse on the Internet, goes on to discuss some implementations try to address this problem by helping users understand the biases of the content they consume, or to consider/explore alternate perspectives, or engage in discourse with a wide variety of viewpoints.

Reflection-

[1] is an interesting read, and makes some fascinating claims, but it has a few flaws. Firstly, it was published around 2009, which was right at the dawn of the age of machine learning for recommender systems. This meant that most websites did not have user specific curated content at that time. The hypothesis discussed by the article, which suggests that the internet may not create echo chambers, since users are not particularly averse to looking at views which go against their own, is not as valid in today’s world. Due to automatic recommender systems, users do not have a choice in this matter anymore, and may be continually exposed to partisan information simply because of their prior information usage patterns.

Secondly, the paper admits that the selection of candidates for the study was not exactly a good representation of the entire nation. The users who signed up for this study already had strong political views, since they were active on either a left leaning or a right leaning website from before. Moreover, more than half of them had a college degree, and the ethnicities of the participants were heavily skewed.

Interestingly, [1] mentions that while only 20,000 people saw the recruitment statement on the left-leaning website (AlterNet), 100,000 people saw the statement on the right-leaning website (WorldNetDaily). However, both sites were almost equally represented in the final selection of candidates, despite the recruitment statement being seen by 5 times as many people on WorldNetDaily. This could hint at an inherent “willingness to participate” of left-leaning people, or might simply be because the readers of the left-leaning site had a lower income on average (as claimed by the paper), and thus desired the participation prize more.

[1] also makes a claim that opinion challenging posts would also lead to an increase in the duration for which the user engages with the content, which is later backed by the data. However, users would probably be less inclined to immediately close articles which they disagreed while interacting with a new unfamiliar software interface, when they know they were taking part in a monitored survey, as they would be when browsing privately.

It is interesting to see that fears of rising political polarization catalyzed by Internet technologies were prevalent not just in 2009, but also as early as 2001, as indicated by the citations made by [1]. It is almost eerie to see these insights become relevant again, more than a decade later.

Many of the systems discussed in [2] would also have a tendency to become biased, depending on the beliefs of the majority share of the users of the systems. For example, if more liberals used Reflect or Opinion Space, then those comments would be more prevalent, and would receive more positive reviews from other liberals.

Opinion Space in [2] reminded me of the abstract interfaces mentioned in the Chat Circles paper, as it creates a space for users to navigate with, where users interact with different types of comments. It also changes the physical characteristics of the comments based on how the users interact with them.

Read More

Papers–

[1] Social Translucence: An Approach to Designing Systems that Support Social Processes – Erickson et. al.

[2] The Chat Circles Series – Explorations in designing abstract graphical communication interfaces – Donath et. al.

Summary –

[1] deals with social translucence in digital networks. It identifies 3 key areas which contribute to providing translucence; visibility, accountability and awareness. It tries to address the fact that certain dimensions of social interactions are not apparent or a given in digital interactions. It focuses mainly on abstract visual representations of the world, as opposed to realist or mimetic approaches. The authors take inspiration from architecture and urban planning, and discuss in great depth a system, babble, which tries to create a socially translucent platform for interaction, describing how it handles conversations, user activity, and knowledge community building.

[2] describes many approaches at building a socially translucent system based on abstract representations of the world and the users, namely Chat Circles, Chat Circles II, Talking in Circles, and Tele-Directions. It goes into more detail about the particular design choices of each system, how they helped ensure social translucency, and observations of users interacting with these systems.

Reflection-

Firstly, these two papers were among the most entertaining and interesting papers I have read in this course so far. There were many interesting observations I could make, including the following:

[1] emphasizes that there is a balance between privacy and openness to facilitate social ease of communication, and knowing these balances is often key to guiding and enforcing social interactions. [1] gives a very interesting example of authors of a book gathering together to finalize the organization of a book, where they used physical space and social rules to their advantage, to create translucency through space and time (i.e. the visibility and the audibility of authors).

Certain platforms enforce translucency through ephemerality, and have mechanisms to prevent users from subverting mechanisms, e.g. Snapchat notifies the user if someone else takes a screenshot of their snaps, thus creating social pressure to enforce ephemerality. [1] talks about Babble, which creates a platform which provides social translucency, but not much information is given about the feedback they received from users about the system, i.e. whether they actually felt more comfortable interacting in this environment, and whether they felt it was natural or forced, and whether it was useful. Moreover, some of the drawbacks of the realist approach highlighted by the authors are not valid anymore. For example,  the limitations on processing speed, number of users, bandwidth, technological support, etc. are not the same as they were 20 years ago.

Interestingly, [1] mentions that for company document databases, people wanted to know which person wrote an entry. Data should be social too, instead of being only dry and descriptive, creating a knowledge community instead of knowledge database. Semiautonomous communities which each aggregate and select information to send to higher up communities, can be democratized too, to encourage privacy. This feature can be successfully implemented in a website which behaves similar to reddit or 4chan.

[1] talked about the need for creations of summaries, indices for conversational data, but there is a need to conserve privacy. However, enough information should be given to new users for them to get a gist of the conversation without them understanding the in-jokes, etc. The anonymity paper dealt with how the system “self-corrected” for this, in the example of 4chan, through certain phrases/lingo or tasks like the triforce. 

[2] highlights the effectiveness of terminology on the users perceptions of other users and the platform, i.e. labelling “lurkers” as listeners gives the active posters the mental image that they are an audience.

Even in a system designed to emulate real world social interactions like Chat Circles, the entire conversation history is stored, to accommodate for the fact that users might be browsing several systems at the same time. This shows that perhaps there are some things inherent to online interactions which no amount of socially conscious design can take away, and that maybe online interaction is a wholly new kind of interaction for which new hybrid systems should be designed. 

An area where [2] lacks is that it is too explicit in its data display. Any sort of user behavior statistic cannot help but be very obvious. E.g., data about user posting makes it very clear which users are “shy” or do not post as much. However, in social interactions, this is not explicitly pointed out, but subtly, almost subconsciously realized. It is a challenge to design a system which allows this to happen.

Encoding into the design the ability for events to serve as icebreakers is also an interesting insight in Chat Circles II. Events in platforms like Facebook are largely crowd fueled, where the users themselves create and consume the content (excluding facebook games), but a platform like Club Penguin/RuneScape on the other hand has events generated by the platform itself, which serve as a common topic of conversation. Also interesting is the fact that people tend to move around in the same area, as an analog to fidgeting, and even do things which are unnatural in the real world, like dancing around each other or forming conga lines, to provide movement based social cues to the conversation. Analyzing the movement patterns of characters in MMORPGs who are idle and simply talking to each other might also be an interesting related project. 

Read More

Papers –

[1] “I always assumed that I wasn’t really that close to [her]”: Reasoning about invisible algorithms in the news feed – Eslami et. al.

[2] Exposure to ideologically diverse news and opinion on Facebook – Bakshy et. al.

Summary –

Read More

Paper-

An army of me: Sockpuppets in online discussion communities. -Kumar,et al.

Summary-

This paper attempts to identify, analyze the behavioral and linguistic patterns of, and find accounts which are  “sockpuppets”, i.e. a group of multiple accounts are controlled by a single user, the puppet master. It collects data from 9 online communities, and conducts an extensive statistical analysis to infer information about the characteristics of sockpuppet accounts, like the way they phrase messages, (e.g. they use I, we, you, etc. more, use shorter comments, have smaller vocabularies, etc.). They prove that the sockpuppet account pairs have similar type of usage, disproving the hypothesis that one account may be like an ordinary user’s account, while the other one is a malicious account. They find that sockpuppets can be supporters (in-agreement), dissenters (disagreeing with each other), or may not be malicious at all. They use a random-forest classifier to predict sock puppets, achieving fairly good results.

Reflection-

Firstly, the study combines multiple very distinct communities together, for data gathering purposes. The communities range from Breitbart news to IGN. We know from previous papers that the behavior depends to some extent on the type of community, so it should be worth examining the differences in the results obtained from community to community. It is interesting to note that the communities themselves have different levels of sockpuppeting, (e.g. Breitbart news has a disproportionately high number, almost 2.5 times that of CNN when adjusting for number of users).

This paper reminds me of work previously discussed in class, especially the paper on anti-social behavior in online communities. This is due to the similar nature of data collection and data driven analysis. This paper has some very interesting ideas to collect statistics or test hypotheses, (e.g. using entropy to convey usage information, and finding which users are non-malignant using Levenshtein distance between the sockpuppet usernames). It has some results similar to the study on anti-social behavior (e.g. sockpuppets make a large number of posts compared to normal users, just like anti-social users). It however makes the interesting find that the readability index (ARI) for sockpuppets is about as high as normal users, as opposed to the same result for found anti-social users.

The study also finds that sockpuppet pairs behave similarly to each other. This brings up the question of what kind of users have more of a tendency to create malicious sockpuppets? Maybe the type of activity and behavior seen in the sockpuppets can be traced to a superset of users (of which puppet masters are only one category), and is a characteristic of that superset. Maybe there are even more similarities with other types of users, like trolls. This is worth investigating.

This paper focuses on pairwise sockpuppeting, and its techniques for finding a sockpuppet group for data collection rely crucially on the IP address. This technique is effective to study “casual” sockpuppeting, where a user is just making multiple accounts to browse different topics or upvote their own answers, but this is fairly harmless when done in an uncoordinated manner by many individual users. These techniques would fail when trying to detect or gather data about co-ordinated, deliberate attacks to propagate misinformation or a personal agenda through sockpuppeting, which is the truly dangerous strain of such behavior. For example, if someone were to hire people to create multiple accounts and spread a false consensus/misinformation, the people doing this could access the website through multiple IP addresses, and conceal the source. It also focuses on pairs of accounts, and not on a huge mass of accounts being controlled by the same user.

Read More

Paper-

Read More