Paper #1: Reverse-engineering censorship in China: Randomized experimentation and participant observation
Paper #2: Algorithmically Bypassing Censorship on Sina Weibo with Nondeterministic Homophone Substitutions

Reflection #1:
This paper tries to fill the knowledge gap on modeling censorship framework in China. The authors perform a randomized experiment on 100 websites owned by both Chinese government and the private sector to find out how the censorship works. They also do interviews with people as well as the censors themselves to get a better idea about the steps of censorship in China. In posts, the authors focus on 4 cases: posts having collective action plan or not, posts for or against the government. The authors tried to control the language, topic, and timing of the posts as much as possible. From the result it seems like there is a 40% prior probability of a post falling under automatic review. Despite this, sites seem to be more reliant on human actions on censorship as their automatic keyword matching systems don’t perform well on separating different posts. The government puts more constraint on the censorship of collective action like protest while all the other types of posts have an equal probability of being censored. The authors tried to account for all edge cases in their study.

Reflection #2:
This paper uses the reverse-engineered knowledge from previous paper to evade the issue of censorship. The paper introduces a non-deterministic (randomized) algorithm using homophones (apparently words sounding very similar). According to their experiment, the homophones are not easily detectable using the automatic algorithm, while robust to understanding by users. From the cost perspective this add additional 15 hour of human labor per homophones. Although this approach seems to be good, China is already known for an abundance of cheap labor. So even if this adds extra cost to the system, it would only work on systems managed by private entities. The authors use of most frequent homophones seems clever. But it depends on how users would react if more posts are censored due to the usage of all possible combination of censoring words. Given that they have already complied with the current state of censorship, I wouldn’t argue against that.

Read More

Paper #1: Uncovering Social Spammers: Social Honeypots + Machine Learning
Paper #2: An Army of Me: Sockpuppets in Online Discussion Communities

Reflection #1:
In this paper authors used honeypots, a type of harmless bots, to collect social spammer information from two early social network sites, Twitter and MySpace. The authors provided a clear motivation for the paper. According to authors email spammers and social spammers are very different. The framework introduced in this paper keeps the precision high for detection of spams. It was deployed in the early stage of the social network sites which probably made it easier to use the specific attributes from the user profile information to automatically detect social spammers. In the recent times, it seems social spammers have already been repurposed for new tasks like introducing misinformation in the network. One of the interesting thing in the paper was using different ensemble based classifier which I didn’t exactly understand. The authors introduced spam precision metric to evaluate their classifier in the wild but didn’t say how it was different from precision metric.

Reflection #2:
Sockpuppets are duplicate accounts to deceive users in a social discussion forum. Authors in this paper looked into 9 online forums to analyze attributes of sockpuppets. The first problem I see with the data collection is using IP addresses to detect sock-puppets. Although authors tried to filter IP addresses that are using NAT, it is not clear if that was effective assuming very few people join discussion forums. The authors did use prior work to characterize other parameters for their framework. Authors found several attributes of the sockpuppets are different from other spammers like bots, trolls etc. Sockpuppets tend to participate in a discussion about controversial topics. They are also more likely to be downvoted. The author used entropy to characterize usage pattern of different sockpuppets but it is not clear if that measurement works. The author used ego network to analyze user-user interaction and found that sockpuppets have higher PageRank in the network. The main success of the paper seems to be finding sockpuppets once one pair has been detected.

Read More

Paper: Linguistic Harbingers of Betrayal: A Case Study on an Online Strategy Game

Summary:
This paper tries to find linguistic cues that can predict “betrayal” in an online game called “Diplomacy”. Diplomacy is an online war strategy with a little different order of actions. Here all the player perform their moves at the same time. This makes it very similar to the problem of “Prisoner’s dilemma”. For this reason, players make and break alliances. The main communication medium for the game is through messages. The authors collected messages between users to find out what type of language cues are used by the users before betraying alliance. The authors found that betrayers express more positive sentiment, politeness while less argumentation and planning in their messages. Based on these attributes authors create a model to predict betrayal which achieves about 57% cross-validation accuracy.

Reflection:
Authors’ choice of Diplomacy was a good source for analyzing the interaction of betrayer and victim. The authors in this paper are overly protective about the effect of time on their relationship which is surprising from the given result when they just ignore the status of the game in predicting betrayal. One of the results of the study that doesn’t make sense is that betrayer’s plan less while the victim’s do more. This is counter-intuitive in the sense if the victim plans more it is likely that they have a better grasp of different situations in the game. This begs the question: what is the effect of different level of experience in the game on betrayal? It is unlikely that a novice player (noob) will betray his alliances. Another thing noticeable in the tables in the paper is the absence of the value of the coefficient for both positive features and negative features in their prediction power. Although this paper provides some interesting insight into the behavior of a betrayer, it doesn’t seem to have any direct application in real life.

Read More

Paper #1: A computational approach to politeness with application to social factors
Paper #2: Language from police body camera footage shows racial disparities in officer respect

Summary #1:
This paper does a qualitative analysis of the linguistic features that relate to politeness. From there the authors create a machine learning model that can be used to automatically measure politeness. Authors use two different websites for testing the generalizability of the model. Based on the result the authors do quantitative analysis on the relationship between politeness and social outcome, the relationship between politeness and power. From the result, it appears that users who are more polite are more likely to be elected as admin and once elected they become less polite.

Summary #2:
This paper looks into police’s interaction with drivers seen from police body-camera which has been adopted recently due to controversy regarding their interaction with the black community. This paper introduces a computational model using only transcription of the speech between an officer and a driver. In the first part of the study, a group of 60 participants was used to rate utterances on two criteria, respect and formality. The study finds that there’s no significant difference in formality from police officers for both white and black community. In case of respect, this is significantly different. Based on these result the authors created a computation model to automatically predict score from the data.

Reflection #1:
This paper tried to infer the relationship between politeness and authority. Thier analysis is in some sense lacking. This becomes more evident from reading the paper 2 which does test for other factors that can affect any inference. For example, in this case, authors don’t check for the factors like the responsibility of admins, age difference etc. Although different levels of moderation capabilities are given to users with different reputation, it is common in StackOverflow for the admins to do a lot of moderation. If you look into the number of duplicate question it will be quite clear that to make the site useful to all types of users strict moderation is necessary for the questions from new users. Another factor that might have an affect on the politeness of the users is the age. It is common in StackOverflow that older users are ruder At the same time, they also have very high rating which correlates with thier chance of being a moderator. In the last election, I think one of the main question was about candidates attitude toward the strictness in moderation (Full disclosure: I have voted for Cody Gray in the last election). So these factors might have some effect on the analysis of politeness.

Reflection #2:
This study was done in an intuitive and simple manner. The authors created a model and tried to find out if it was affected by any control variables like severity of the offence, formality, outlier in data etc. The first thing that comes to mind from the method is that the model only focuses on the utterance in textual format rather than speech. It doesn’t appear to be a good ground truth as the RMSE of average users is about .842. The authors uses this partial ground truth from the rating of the human raters to build a computational model for predicting respectfulness in utterances. Another problem with the computational model is the transcribing part of the data is fully manual. In that perspective this is a semi-automatic model. A complex approach will be by using speech directly which solves the previous problem.

Read More

Paper #1: Echo Chambers Online?: Politically Motivated Selective Exposure among Internet News Users
Paper #2: Exposure to ideologically diverse news and opinion on Facebook

Summary #1:
This paper checks whether selective exposure plays any significant role in people’s choice of a news article in high-choice information environment like the Internet. The traditional approach of opinion reinforcement seeking and opinion challenge avoidance doesn’t make sense on the Internet. This paper suggests that this is due to people’s tendency to more focus on opinion reinforcement than filtering opinion challenging information. The study was run on 700+ people recruited online. Each of the participants is shown a set of articles related to a politically relevant topic and asked their interest in reading those articles in two phase. Thier reading time of the articles is also recorded. Based on the result it appears that people tend to select more stories with opinion reinforcement while spending more time on opinion challenging articles.

Summary #2:
This paper compares the influence of algorithm and user’s clickthrough in exposure to cross-cutting stories on Facebook. The network structure in Facebook is different than other social media due to connections in different offline context (school, family, social activity etc.). This is probably why homophily structure in Facebook is different. The authors use 10 million Facebook users who self-reported their political affiliation and their click activity on 7 million web URLs shared to analyze the influence. The analysis was limited to hard contents that have political significance. User’s news feed algorithm depends on various things on Facebook including how many times they visit the site daily, how much they interact with their friends, how often they click certain web URLs etc. The authors’ found that algorithmic exposure to cross-cutting content for conservative users was significantly higher than liberal users while the random probability of seeing cross-cutting content was the opposite. The click rate further limits their exposure.

Reflection #1:
This study was designed in a very simplistic way. The author has already noted that population was selected from two partisan sites which might have influenced the result. By selecting a different set of users it is easy to check if people with more partisan view are more likely to want to see the opposing view. Author’s inference of the non-existence of echo chamber is a bit of a stretch. One thing interesting in the result was that people had more interest in “gay marriage” than “civil liberty” while their reading time on “gay marriage” is less than that of “civil liberty”. This conflicts with the common notion that people spend more time things of interest. People seem to find most stories more opinion challenging than opinion reinforcing. The variance in this opinion seems to increase when people have read the stories. Another variation of this could be checking how much difference it makes in case of just headlines of a story is presented to users. The study didn’t ask if the users already read a story. This could make a significant difference given that news aggregator tend to show most popular stories first. This could also account for less time people spend on stories of interest. One thing I didn’t understand was why use log of the time to fit a linear regression model as the time span was in the range of 4sec – 9min. Another particular odd inference was author thought the correlation of reading time with Age was not significant with coefficient .01 for p < .01 while it was significant in case of opinion reinforcement (retrospective) with coefficient .02 with p < .05.

Reflection #2:
This paper checks what is the role of users click through in the limiting exposure to diverse content. Although the user selection process had a self-reporting bias, the proportion of liberal and conservative users were very close surprisingly. Figure 3 has several implications given certain condition is met. They failed to mention the number of average shares by both conservative and liberal users. Although they mentioned the total proportion of different types of shares by both conservative and liberal users, the average number of shares could skew the result. Similar relation could be found in the proportion of liberal and conservative contents in the 226,000 hard contents. They also didn’t mention users’ living area (city or rural) which could skew the percentage of ties. The authors did find a correlation between the position of a story in the news feed and click through ratio which might have affected the inference. They used a technique to calculate the risk ratio (results not shown in the supplementary text) but provided a partial method for calculation.

Read More

Paper: Antisocial Behavior in Online Discussion Communities

Summary:
Although antisocial behavior in online communities is very common, most of the recent research on this subject has focused on qualitative analysis using a small group of users. This paper uses data from three online communities (CNN.com, breitbart.com, and IGN.com) for quantitative analysis to get a general understanding of the antisocial behavior of users. For the sake of comparison, the paper discusses two types of users: Future-Banned-Users (FBU) and Never-Banned-Users (NBU). By analyzing the post throughout their activity span on the forum authors find that posts by the FBUs are very different than other posts in the thread and harder to read. They also find that their quality of posts worsens over time. Censorship also plays a role in the guiding the writing style of the FBUs. FBUs post a lot and get a higher number of responses. Another finding of the study is that two types of FBUs exist in an online community: one with higher post deletion rate and other with lower. Finally, the authors use four types of features to create a classifier for predicting antisocial behavior: post feature, activity feature, community feature and moderator feature.

Reflection:
The paper explains the process of analyzing antisocial behavior starting from data preparation to analysis in great detail. One interesting aspect of the process was using crowdsourcing for initial classification. The authors’ analysis of the final classifier provided some interesting insight. For example, classifiers performance peaks on seeing the attributes from first 10 posts. This correlates with the idea that other community members judge FBUs in a similar fashion. The performance of the classifier on Hi-FBUs suggests that the classifier learns the post deletion ratio as one of the primary indicators which explains why prediction performance peaks at seeing first 10 posts. The authors’ analysis of the cross-platform performance of the algorithm was very intuitive. Although the prediction quality of the classifier is good enough, there remains the issue of application of such tools. Finally, this paper discusses a sensitive issue of antisocial behavior and creates a tool for prediction. Although the performance is good enough, still there is a necessity of the human factor in preventing such behavior.

Read More

Paper:- The Language that Gets People to Give: Phrases that Predict Success on Kickstarter

Summary:
This paper looks into the language usage and its predictive power of getting funded in a Crowdfunding site, Kickstarter. Previous studies already found several factors that affect funding probability in such sites, like higher funding goal, longer project duration, video in a project pitch, social network, key attributes of the project etc. This study builds on that by adding linguistic analysis of the project pitch. The authors apply the unigram, bigram, and trigram phrases common in all 13 categories of projects as linguistic predictive variables along with 59 other Kickstarter variable into a penalized logistic regression classifier. Finally, authors do both qualitative and quantitative analysis of the result of the output. Authors provide top 100 phrases that contributed to their algorithm. From the qualitative analysis, several intuitive phenomena appear like reciprocity and scarcity have positive correlation with being funded. Several other factors like social identity, social proof, and authority also seem to contribute to the process. LIWC analysis suggests that funded project pitch includes higher cognitive process, social process, perception rates etc. Although sentiment analysis shows that funded projects have higher positive and negative sentiment, it is not statistically significant. One interesting phenomenon found in the analysis was that a completely new project is likely to have less success than one that builds on a previous one.

Reflection:
The paper provides a good motivation for the analysis of linguistic features in Crowdfunding projects. The authors use of Penalized logistic regression seemed interesting to me. I would have probably thought of applying PCA first and then doing a logistic regression. But Penalized logistic regression provided results which were important for interpretation. At the same time, looking into the top 100 positively and negatively correlated terms reminds about the fault of big data interpretation: “seeing correlation where none exists”. For example, “Christina”,”cats” etc. For the sake of generalizability, authors lose many terms which could have better correlation with projects. But the beta score of the 29 control variables says otherwise. Authors’ use of Google’s 1T corpus for reducing the number of phrases and tree visualization of some common terms were nice additions to the paper. Although positively correlated terms don’t guarantee success in a Crowdfunding project, negatively correlated terms provide a list of things to avoid in the project pitch which is very useful. The social proof attribute of the result begs the question, can we manipulate the system by faking backers?

Read More

Paper #1: Critical questions for big data: Provocations for a cultural, technological, and scholarly phenomenon
Paper #2: Data ex Machina: Introduction to Big Data

Summary #1:
This paper discusses what is entailed by the phenomenon of big data from the perspective of socio-technical research. The authors describe Big Data phenomenon as an interplay between technology, analysis, and mythology. Here mythology represents the uncertainty of inferring knowledge from big data. The theme of the paper is six provocations about the issues in big data: knowledge definition, objectivity claim, methodological issue, importance of context, ethics, and digital divide due to access restriction. Big data introduces new tools and research methods for information transformation which in turn changes the definition of knowledge. While quantifying information in big data might seem to provide objective claim, it is still the subjective observation that initiates it. It’s the design decisions in the interpretation of data that incorporate subjectivity and its biases. One caveat of interpreting big data is that seeing correlation where none exists (p.11). The methodological limitation of big data lies in the source and collection procedure. It is necessary to understand how much the data was filtered and if it is generalizable to the research domain. One solution to this problem is using heterogeneous sources to combine information which also amplifies noise. This heterogeneous combination is also helpful in modeling the big picture. For example, Facebook can alone provide a relationship graph of a user. But it is not necessarily the overall representation. Because multi-dimensional communication methods like social network, email, phone etc. each provide a new representation. So context is very important in describing such information. Big data also raises the question of ethics in research. The sheer volume of big data could provide enough information to de-anonymize. Information should be carefully published to protect the privacy of the entities involved. Finally, accessibility of big data divides the research community into two groups where one group is more privileged. Computational background also creates similar division among big data researchers.

Summary #2:
This paper is similar to the previous one in the sense it discusses similar issues from the previous paper. The authors here first discusses the data sources for big data by dividing them into three topological categories: digital life, digital trace, digitalized life. Digital life refers to the digitally mediated interactions like tweeting, searching etc. Digital traces are the records that indirectly provides information about the digital life like Call detail records. Finally digitalized life represent the capture of a nondigital portion of life into a digital form like constant video recording in an ATM. There is also possibility of collecting specific behavior like certain types of tweets or visiting certain webpage. These data provide several opportunities for behavioral research. Big data provides large data set from different sources and combination of these sources provides important incites like the Copenhagen Network Study. Big data also provides cost-effective solutions to some studies like unemployment detection, disease propagation study etc. Effect of external changes can be captured by big data like Hurricane, price hike etc. By focusing on underrepresented population, big data is used to study certain problems like PTSD, suicidal ideation etc. The vulnerabilities of big data include the problem of generalizability of hypothesis, heterogeneous sources, errors in the source systems, and ideal user assumption. Research on big data includes ethical issues in both acquisition and publishing of data. Finally, recent big data trends include: new sources of data, generic model for research, qualitative approach in big data research etc.

Reflection:
Both of the paper discusses issues and application of big data in identifying social phenomena. This reflection focuses on the generalizability issue in big data. The authors suggest combining multiple sources to validate can solve generalizability issue. This seems interesting given recent deep learning community has found that generalizing a model can be achieved using more data as well as using transfer learning. Similar approach can be used in finding social phenomenon in big data. For example, data from Twitter can provide with information about the spreading of rumors by people with certain attributes. Although Facebook is too different from Twitter, it is possible to use the hypothesis and the result from Twitter to initiate a learning model to apply in facebook. What do you think?

Read More