Thursday, July 6, 2017

2017-07-06: Web Science 2017 Trip Report

I was fortunate enough to have the opportunity to present Yasmin AlNoamany's work at Web Science 2017. Dr. Nelson offers an excellent class on Web Science, but it has been years since I had taken it and I still was uncertain about the current state of the art.
Web Science 2017 took place in Troy, a small city in upstate New York that is home to Rensselaer Polytechnic Institute (RPI). The RPI team had organized an excellent conference focused on a variety of Web Science topics, including cyber bullying, taxonomies, social media, and ethics.

Keynote Speakers


Day One


The opening keynote by Steffen Staab from the Institute for Web Science and Technologies (WeST) was entitled "The Web We Want". He discussed how we need to determine what values we want to meet before deciding on the web we want. Dr. Staab defined three key values: accessibility for the disabled, freedom from harassment, and a useful semantic web.

Staab detailed the MAMEM project whose goal is to provide access to the web for the disabled, accounting for those without the ability to operate a mouse and keyboard as well as those who cannot see or hear. He mentioned that the z-stacking used by the Google search engine's textbox frustrates a lot of accessibility tools.


On the topic of harassment, Staab indicated that we need to determine the roles and structure used by people in social networks. Who is each person linked to? Are they initiators or do they join discussions later? Are they trolls? Are they contributors? Are they moderators? Can we differentiate these roles based on previous experience? He showed the procedure by which the ROBUST project classifies users into each of these roles with the goal of providing an early response to trolls, attacks, and spam.




For a useful Semantic Web, Staab stressed the importance of data that interlinked and allows us to further describe entities. Most quality assessments for existing links don't take into account the usefulness of the data. How close are we to benchmarking this usefulness? It depends on the application. So far, we have recommender systems that work based on what someone else said was useful, but may not fit the needs of the user under consideration. Tests of usefulness are further frustrated by the fact that people behave differently during testing the second time through.

He closed by stressing that we need to measure our achievement of these goals. As we measure the achievement of these values, it inspires our engineering and this same measurement is required to understand how well an engineering solution works.

Day Two


The second day keynote was by Jennifer Golbeck, world leader in social media research and science communication, creator of the field of social analytics, and professor at the University of Maryland. She started by talking about one of the reviewers of her paper, "A Large Labeled Corpus for Online Harassment Research", submitted to Web Science 2017. This reviewer liked the work done on social media harassment, but objected to the inclusion of harassing tweets as evidence in the paper. Dr. Golbeck objected to this idea that we should not include evidence in scientific papers, no matter how offensive it may be. She used the story of the upset reviewer throughout the rest of her talk.
Her research tries to answer questions such as who is posting harassing content and why? She also mentioned that Twitter will often not help you block someone if you report them. In order to study the phenomenon, she sought out harassing tweets on Twitter. Fortunately, there is a low density of harassing tweets in the Twitter firehouse. After several unsuccessful methods, including blocklists like Block Together, she resorted to finding harassing tweets using Twitter searches by combining expletives and the names of marginalized groups. Harassment is directed at these groups because they have less power and it is intended to silence them. Sadly, 50% of the tweets containing the word "feminist" are part of harassment on Twitter.


She discovered that there were several main groups of harassers in the data, labeled Gamergate, Trump Supporters, Alt-right, UK-based Brexit/anti-Muslim, and "Anime". This does not mean that all Trump Supporters harass people, but there is a large group of harassers that are Trump Supporters. The "Anime" group seemed to be very interested in the Japanese cartoon style, but not all Anime fans are trolls, and likewise with other descriptors.


She also highlighted the work "Trolls Just Want to Have Fun", by Buckels, Trapnell, and Paulhus. Buckels discovered that trolls exhibit a higher percentage of the following personality traits: machiavellianism, narcissism, psychopathy, and sadism. This was contrasted with those who merely engage in debating issues. These personality traits are known in psychological circles as the Dark Tetrad of personality, identifying individuals that are more likely to cause "social distress".


In spite of some of this progress, we still have no idea of the full picture of harassment on Twitter. One would need to learn the language of the communities under study, both of harassers and victims, in order to fully discover all of the harassment going on.  This makes members of these groups -- women, minorities, etc. -- more careful about what they say on social media because they have to weigh the potential harassment before even speaking. She returned to the reviewer of her paper and stated that she included the Tweets not only as evidence, but because the more we are silent on these issues, the more they will continue.

I Presented Yasmin AlNoamany's Work


On the third day, I was fortunate enough to present Dr. AlNoamany's work on using storytelling tools to summarize Archive-It collections. She uses the example of the Egyptian Revolution, much of which was recorded online in real time, as a use case for summarizing web archive collections. Much of the web resources from the Revolution are gone, but have been preserved in web archives.


csvconfyasmin2017_05_03 from Yasmin AlNoamany, PhD

There are multiple archive collections about the revolution and it is difficult to visualize more than potentially 1000 different captures of potentially 1000 seed URIs. We seek to answer questions such as: "What is in this collection?" and "What makes this collection different from another?" She uses social media Storytelling as an existing interface with which users are familiar. This presentation discusses, at a high level, the Dark and Stormy Archives (DSA) framework which automatically summarizes the collection and generates the visualization in Storify.

Selected Posters


There were many excellent posters at Web Science 2017. Unfortunately, I do not have room to cover them all, so I will highlight a select few.


In "Understanding Parasocial Breakups on Twitter" (preprint), Kiran Garimella studied perceived virtual relationships, known as para-social relationships, on Twitter. This scenario erupts when a user follows a celebrity on social media and then are followed back. For some fans, this provides the illusion of a real relationship. A para-social breakup (PSB) occurs when a fan stops following the celebrity. He studied the 15 most followed celebrities from popular culture on Twitter and used a subset of their fans. He classified fans into 3 types: (1) involved, (2) casual, and (3) random. The involved fans tweet often with their chosen celebrity, but also have a higher probability of unfollowing the celebrity than casual fans who tweet with their celebrity only once per year, or a random sample of followers. Garimella's study has implications for marketing.

The mobile game Pokémon Go has a feature known as the Pokéstop where players can gather more resources to continue playing the game. In "Pokémon Go: Impact on Yelp Restaurant Reviews" (preprint), Pavan Kondamundi evaluated whether or not the inclusion of Pokéstops in Yelp restaurant profiles had an impact on the reviews for said restaurants.  His study included 100 restaurants, half of which contained Pokéstops. He found an increase in the number of reviews for the period of 2014 to 2015, but a slight decrease for the period of 2016-2016.

Policy documents are used by many organizations, not just those within government. Bharat Kale's work, "Predicting Research that will be cited in Policy Documents" (preprint), attempts to determine what features increase the probability of an academic work being cited by a policy document. Using features related to online attention, he discovered that the Random Forest classifier showed the best results for predicting if an article is cited by a policy document. Mention counts on peer-review platforms, such as PubPeer, seems to be the most influential feature and mentions in news appears to be the least influential feature. He intends to extend the work "to predict the number of policy citations a given work is likely to receive."
As I mentioned in an earlier blog post, the problem of entity resolution, and more specifically author disambiguation, continues to confound solutions for scholarly communication. Janaína Gomide focuses on the synonym problem, where a single individual has multiple names. In her work "Consolidating identities of authors through egonet structure", rather than using content information about a given author, she is studying egonets, networks of collaborators built from co-authorship information. She is developing an algorithm that attempts to disambiguate authors based on the shape of their egonet. Preliminary results with datasets from DBLP and Google Scholar show promise for the current version of this algorithm.

There was a lot of work on social networks at the Web Science conference, and Nirmal Sivaraman's work "On Social Synchrony in Online Social Networks" was no exception. He defines social synchrony as "a kind of collective social activity associated with some event where the number of people who participate in the activity first increases exponentially and then decreases exponentially". He developed an algorithm that determines if synchrony has occurred within a dataset of social media data.

Spencer Norris won the best poster award for his "A Semantic Workflow Approach to Web Science Analytics". He highlights the use of linked data to build workflows for use in running and repeating scientific experiments.  His work focuses on the use of semantic workflows for Web Science, indicating that these workflows, because of their ease of publication and analysis, also easily allow "Web Science analyses to be recreated and recombined". He combines the Workflow INstance Generation and Specialization (WINGS) system with the existing Semantic Numeric Exploration Technology (SemNExT) framework.

Selected Papers

There were 45 papers accepted at Web Science. I will summarize a few here to convey the type of research being conducted at the conference.

The design of web pages have shifted over time, leading to differences in how we consume them. Bardia Doosti presented "A Deep Study into the History of Web Design" (copy on author's website). In their work, they point out that web design, much like paintings and architecture, can be analyzed to indicate the concepts and ideas that represent the era from which a web page comes. They develop several automated techniques for analyzing archived web pages, including the use of deep Convolutional Neural Networks, with a hope of identifying the web pages' subject areas as well as determining which web sites (such as apple.com) may influence the design of others.

Olga Zagovora presented "The gendered presentation of professions on Wikipedia" (preprint) where she and her co-authors conducted a study comparing the number of women mentioned on the profession pages of German Wikipedia to the actual number of women in those professions, indicating that there is still a gender bias in the pages. They compared the number of images, mentioned persons, and wiki page titles). It is likely that the choices representing individuals in these professions may be made out of tradition or due to the historical preponderance of males in these fields, but this work is useful in informing further development of guidelines for the Wikipedia community. The data is available on GitHub.

Companies, celebrities, and even general users want to know what helps them acquire more Twitter followers. Juergen Mueller and his co-authors attempted to determine the influence of a user's profile information on her number of followers in "Predicting Rising Follower Counts on Twitter Using Profile Information" (preprint). Because of the rate limitations of the Twitter API, they are interested in determining what information can be predicted based on a user's profile. Using several classifiers, they discovered that follower account is affected by the "subjective impression" of the profile's name, indicating that follower counts are adversely affected for accounts with a name that is perceived as feminine. They also confirmed earlier research that indicates that users with a given name in their name field have less followers.


The concept of fake news has a lot of media attention, especially since the 2016 US Presidential Elections. In "The Fake News Spreading Plague: Was it Preventable?" (preprint), Eni Mustafaraj presented recipes for spreading misinformation on Twitter from 2010 and spreading fake news on Facebook from 2016. The two recipes have the same steps, meaning that perhaps the spread of misinformation during the 2016 US Presidential elections could have been avoided. She mentioned that even thought Facebook had been working on preventing the spread of hoaxes since January of 2015, they were unsuccessful.


Omar Alonso and his colleagues at Microsoft created a search engine in "What's Happening and What Happened: Searching the Social Web". They are building a growing archive of tweets to find relevant links that had been shared on social media. Their project differs from others because they are also adding a temporal dimension to their data gathering to show what people were talking about at a given time, which keeps the search engine fresh, but also allows for some historical analysis. The system uses a concept of virality rather than just popularity for the inclusion of results. Because of this focus on virality, their system is able filter fake news from the results. Contrary to other results, they discovered that "the total number of shares of the real links was higher than the fake links" on Twitter. The resulting search engine allows a user to search for a topic at a given date and time and discover what links were relevant to that topic at that time. The results are presented as a series of social cards rather than the "10 blue links" presented by well known web search engines. These social cards are similar to the link cards used in Storify: they contain an image, title, and short description of the link behind the card.



Alonso also presented the work "Automatic Generation of Timelines from Social Data", which attempts to determine what occurred on a given day for a given hashtag. The system evaluates the tweets by several metrics for relevance, quality, and popularity to produce a vector of relevant n-grams for that hashtag. Once this is done, links are extracted from the tweets, and titles are extracted from these links. The document link titles are evaluated using a new technique the authors name Social Pseudo Relevance Feedback which combines their existing n-gram vectors with the concept of pseudo relevance feedback from information retrieval in order to re-rank the link titles. The highest ranked title for the time period, a day or an hour, is then presented as an entry into the story. The dates can then be listed next to the title produced for that date which, when presented in this fashion, represents a timeline of events matching a given hashtag (seen for #parisattacks and #deflategate in the photos above). I thought this was quite brilliant. One could easily extend this technique by presenting the generated links in order using a tool like Storify, much like Dr. AlNoamany has done for Archive-It collections.


Web Archives are important to the research of the Old Dominion University WS-DL group, so I was intrigued by "Observing Web Archives" (university repository version) presented by Jessica Ogden. She was interested in the in-depth "day-to-day decisions, activities and processes that facilitate web archiving in practice". She used an ethnographic approach to understand the practices at the Internet Archive.





Kiran Garimella presented "The Effect of Collective Attention on Controversial Debates on Social Media" (preprint), studying polarized debates on Twitter. They analyzed four controversial topics on Twitter for from 2011 - 2016. They discovered that "spikes in interest correspond to an increase in the controversy of the discussion" and that these spikes result in changes to the vocabulary being used during the discussions as each side argues their case. They want to develop a model that allows us to use "early signals" from social media to "predict the impact of an event". Kiran won best student paper for their study.

As Web Science researchers, we spend a lot of time analyzing the data available from the web. While working on the online harassment Digital Wildfire project, Helena Webb, Marina Jirotka, and their co-authors began to question the ethics of exploring a user's Twitter data without a user's consent. Even though Twitter is largely a public social network, it presents issues when one considers that researchers are deriving behaviors and information about people and thus has parallels with the testing of human subjects. Marina Jirotka presented "The Ethical Challenges of Publishing Twitter Data for Research Dissemination" (link to university repository). They indicated that there is indeed harm to be caused by republishing social media posts, exposing the attacker to retaliation and forcing the victim to relive the experience. Even if one were to anonymize posts, it is still difficult to fully anonymize the subject, considering the posts can be found via search engines on social media sites.

If a researcher wanted to acquire consent, how would they do so? In the case of Twitter, the social media feed is so large that many users do not view it all and may miss requests for consent. How often should the researcher attempt to contact them? Is an opt-out policy better than an opt-in policy? Echoing Jennifer Golbeck's keynote: If posts were observed, but cannot be included in published research, how do we support our findings? The study exposes many of these concerns in hopes that we can come to a consensus on how to handle them as a community. This study won best paper.


Everything Else








There was a panel discussion at the end of the first day on the ethics of web science. It echoed some of the issues brought up in Helena Webb and Marina Jirotka's paper, but also introduced some additional perspectives. The panel consisted of Jim Hendler, Jeanna Matthews, Steffen Staab, and Hans Akkermans. Each offered many different perspectives, but it was clear that the concern is that Web Science researchers need to drive the ethics discussion before groups outside of the community drive it for them.



Of course, we enjoyed the time learning from one another. Discussions over dinner were influenced by the presentations we had witnessed during the day. We were also able to educate one another about our individual projects. Memento made an appearance in some of those discussions and stickers ended up on some laptops.


I would like to thank John Erickson, Juergen Mueller, Lee Fiorio, Jim Hendler, Omar Alonso, Wendy Hall, Kiran Garimella, Jessica Ogden, Marina Jirotka, James McCusker, Deborah McGuinness, Olga Zagovora, Frederick Ayala-Gómez, Hamed Alhoori, Peter Fox, Spencer Norris, Katharina Kinder-Kurlanda, Eni Mustafaraj, Xiaogang (Marshall) Ma, and many others for fascinating insight and interesting discussions over meals and outings.

Summary



The Web Science 2017 conference was invigorating and fascinating. It has really inspired me to make Web Science an area of interest in future studies. The Web Science Trust has summarized the conference and also provided a Storify story of what happened. I am looking forward to possibly attending this conference again in Amsterdam in 2018 where I may contribute the next grains of knowledge to the discipline.

-- Shawn M. Jones

No comments:

Post a Comment