05/07: State

[shevajia]

Comment below with questions or thoughts about the reading for this week's workshop.

Please make your comments by Wednesday 11:59 PM, and upvote at least five of your peers' comments on Thursday prior to the workshop. You need to use 'thumbs-up' for your reactions to count towards 'top comments,' but you can use other emojis on top of the thumbs up.

[wanitchayap]

Thank you for sharing and presenting your work. This is a very well designed and controlled study! Below are some questions I have:

1. Using FB friends as a proxy of the social network - Do you think individual differences in adding FB friends could confound the results? For example, some people may add any suggested friends of existed friends without actually knowing them and they might never interact with each other even after being friends on FB. On the other hand, some people only add FB friends that they are actual friends in real life. How would you account for such individual differences? Can we use longitudinal data of the user's behavior in adding FB friends (does a particular user typically add friends from FB suggestion or never at all?) Should the study takes into account the degree of activities shared and exchanged between a pair of FB friends (to avoid measuring networks that actually lead to no connection benefit in the future)?

2. I am quite surprised about the t-SNE result on the Minority−serving feature. Do you think this lack of pattern stems from the actual truth or actually from t-SNE being very flexible? You mention that the different hyperparameters give similar visual results. However, do you think this fact alone is enough to confidently say there is no pattern (except for HBCU)? In addition, is there any interpretation of the shape of the t-SNE (that is with a big hole on the upper right corner and some small holes on the upper left)? I am personally a bit skeptical about t-SNE. Do you think there is a better way to visualize or extract the results besides t-SNE?

[bakerwho]

Thanks for sharing your work! My question is about your work at Facebook. Can you please tell us a little more about how you have improved the performance of ranking models? Are these ranking models for pages and people? Does this have any bearing on the prioritization of content shown in people's feeds?

I'm particularly interested in Facebook's News Feed algorithm, which has gotten a lot of recent attention for its power in deciding what types of information people see about the world. I've read a little about your work with Paolo Parigi on trust-engineering and would love to hear about it. I'm curious about Facebook's approach to trust, rankings, and content-prioritization.

[liu431]

Thank you for sharing the talk. I am thinking about how can this analysis and data create value for Facebook users (in particular, 12th-grade students who are contemplating which college to choose). Currently, there are many college rankings based on social data, such as the ones by Linkedin and Niche. However, I found them not very useful when I was choosing colleges. Thus, it would be interesting if Facebook could provide a personalized college recommendation algorithm using this data. For example, for a nerdy student who don't like much social life, the algorithm should easily recommend UChicago over other similar colleges for them.

I am wondering if this sounds useful or interesting for your team to implement in the future?

[rkcatipon]

Thank you for sharing your work! I'm really interested in the social formations caused by schooling. Often, private schools are considered more valuable, not just because of the educations provided, but because of the socio-economic status of their populations. Education is this sense can be seen as a class codifying mechanism. Towards the conclusion, you mention some of the factors that contribute to a tie formation such as socio-economic and cultural backgrounds. I would be really interested in looking at the socio-economic and other demographic characteristics of these networks to see if like groups tend to cluster together or if schools are a way to bridge these differences!

Separately, I have a question about user demographics. In recent years, young people have moved away from Facebook to other platforms like Instagram, Snapchat, and TikTok. While they may maintain their accounts, it is unclear if Facebook is as relevant to Gen Z'ers as it is to other generations. Considering this change in user behavior, do you think the shape of networks generated by the pairwise ego-graphs may shift in response?

[anqi-hu]

Thank you for sharing your work with us! It is fascinating to see school-wide networks analyzed by school characteristics. I'm wondering if calculating degrees of separation for each college/university might provide insight on how well the Facebook friendship network is representative of the real-world case and what school characteristics is related to varying degrees of closeness. Do you think it is meaningful in this context to see if the clustering observed at these institutions also mean high connectedness throughout the cohorts, or if they actually mean more sparsely located, smaller clusters of connections?

[WMhYang]

Thank you very much for sharing this interesting work. For me this result would be extremely useful if I were to choose a university to continue my study. I am considering the value of the networks formed in the university.

Firstly, as mentioned by @rkcatipon, socio-economic status would be important in the networks, so I am wondering whether schools could smooth the difference in socio-economics status and what school level differences could account for the difference in socio-economic status.

Secondly, I think that the friends whom you always contact would be more valuable to you, so I am thinking of an extension but I am not sure whether it is feasible/valuable. Could we weight the edges by the frequency of messages between two individuals so that we can focus more on the tight connections instead of the loose connections?

[adarshmathew]

Thank you for sharing your work with us! This is a really rich dataset, and I'm excited to see how it'll be used in the future.

Using this dataset to quantify the impact of networking on college earnings I'm curious about how researchers can use this rich dataset to quantify the 'intangibles' of college (especially salient in these times of remote learning). Ballooning tuition costs and student debt are a couple of reasons why the DoE created the College Scorecard. If I remember correctly, they plan on including average earnings and debt for academic programs, along with the yearly tuition in future updates -- essentially creating a metric to see if some programs give you better value for money. University administrators have pushed back against this framing, regularly arguing that a college education is so much more than the sum of the courses you take and your immediate earnings.

What do you think of the possibility of using your dataset to explain that some of the relative 'efficiency' of certain programs comes from the networks students form in their time there? There's a rich literature which talks about the importance of networks and prestige of college in future earnings of graduates. Being able to quantify the 'networking effect' would help students and administrators measure the worth of cross-disciplinary academic programs and professional events.

Distribution of AUC values across all pairs Your construction of the 'similarity' metric built on AUC of a Random Forest classifier is very interesting, and makes a lot of sense for the identification of similarity. And your collection of examples across types of universities and their characteristics sheds some suggestive light about what could explain the dynamics of network formation at these various schools. But could you tell us about the distribution of your AUC values across all 7660C2 pairs of entry cohorts you look at? My intuition -- and I may be completely wrong about this -- suggests that we'd see more difficult-to-distinguish pairs than not (count of low AUC pairs >> count of high AUC pairs). Because your ego-sampling technique, at its heart, captures the individual-level dynamic of friendship formation in college, which may be fairly stable across schools.

Bonus questions:

• Why the choice of 250 for the ego sub-samples? Do you think your analysis would be robust to perturbations to this number?
• Are your feature importance distributions all unimodal? I would love to see a version of Figure 5 that's a pairwise scatterplot of all the features and their importance values. I'd be curious to examine the subsset of cases where, for example, eigenvector centrality had low importance, and which other features were deemed more important.
• Are there any alternatives to the AUC-based measure to similarity, something topological perhaps?

[ShuyanHuang]

Thank you for sharing your work with us. I think it would be more interesting if any time-varying properties of the network structures can be studied. For example, how does similarity of the college cohort network evolve from year 1 to year 4?

[lulululugagaga]

Thank you for your sharing, I think this work would be especially interesting when we're all in quarantine. My question is that social networks may be broad, when you get the data and start to explore this topic, do you feel like it may turn out to be a data exploratory analysis instead of sticking to a very focused question? I sometimes lose consistency.

[vinsonyz]

Thank you for sharing. My question is that is there any way that students can access such a dataset of Facebook users? One main concern when we think of some research questions is that we do not have access to proper data.

[nwrim]

Welcome to our workshop! Based on personal experiences, I feel like the dominance of Facebook in college networking space diminished as time went - by the introduction of alternatives and etc. Were there any decrease in the size of networks (both by the number of participants or by the number of connections) as the time-series advances? If so, what could be a way to incorporate this change?

[dhruvalb]

Thank you for sharing this work! This work was limited to US institutions. Is there any similar work for to understand characteristics of universities in other countries? Would you hypothesize similar findings?

Just to clarify, it is noted that the focus is "undergraduate 4-year higher education establishments" but is it limited to undergraduate students or does this data capture grads and undergrads at such institutions?

[bhargavvader]

I want to +1 @vinsonyz . How is Facebook working towards making this data more accessible? Prof. King gave a seminar here about anonymising FB data, and we all know how much FB needs to keep their data private because their entire business is making intrusive models to model our behaviour and then sell intrusive annoying advertisements. How do you personally reconcile working for a company with such gross privacy violations with doing cutting edge research? I like cool research too but I guess I just don't like big corporations and the way they will do anything to keep their revenues going.

[ydeng117]

Thank you for sharing this work. First, I wonder how does the population of international students can affect the social network. Do students from the same country tend to add each other? What are the ways for international students to build up networks with native students? My second question is that how subcultural clubs other than fraternity and sorority affect the networks?

[zeyuxu1997]

Thanks for sharing your interesting work. I have some questions about how to construct the dataset. You used the data from Facebook to study the social network of students, have you considered to find other data sources? For example, I think LinkedIn may be a popular social platform and contains resumes of users. Could you please introduce some methods and tips to construct a dataset with data from different sources?

[luxin-tian]

Thank you for sharing your work! It's a fascinating research into social networks at a college level. How can data science help further dig into the mechanisms that form the structure?

[dongchengecon]

Thank you so much for your presentation! The topic of college network structure is quite interesting. As for the possible extension of your research, I am thinking about creating social connection measurements between schools. Generally, you could use the ratio of the number of connections on Facebook between two schools with respect to the multiplied number of users to measure connection. Besides, is there any possibility that the data could be used to test the theory on network formation process?

[ziwnchen]

Thanks for sharing this interesting work! Despite the inevitable limitation/bias because of using Facebook data as a proxy of the college network mentioned by @wanitchayap @rkcatipon and the paper, this research is still fascinating because it is the first time the college network could be empirically examined with large scale data! But still, I am curious about the generalizability of findings in this research: the article does mention that FB data has some degree of representativeness, but it is still unclear what kind of population this finding could be applied to. Another thing is, could you briefly tell us about the logic behind the multiple machine learning methods you used here?

[tonofshell]

I think this is a very interesting use of Facebook data, but I wonder what being Facebook friends actually captures about online or in-person friendships and connections. At least from my personal experience, it is very rare that anyone unfriends someone unless there has been a very serious falling-out. Therefore, especially during college, you often accumulate many Facebook friends that you have very little to no interaction with on Facebook, much less offline. Did you look at any data about how much people interacted on Facebook? Do you think your results would differ when looking at only "active friendships" instead of all Facebook friendships?

[linghui-wu]

This is such an interesting topic and I really look forward to it! My question is similar to the first point of @WMhYang. I have encountered relevant papers in economics on social mobility before and curious about how network analysis can provide new insights for this topic.

[di-Tong]

Thanks for sharing this interesting paper! I wonder if it is possible to test whether the networks and their patterns you identified here persist throughout people's lives (which you argued in stressing the implication/importance of this research).

[KenChenCompEcon]

Thanks so much for presenting. This is a very interesting topic. I am personally curious about do some schools share similar patterns of networks than other school pairs? Is it possible to also reveal the connection between different schools leveraging similar methodology? This might help explain networking out of school and in workplaces.

[YuxinNg]

Thank you for the presentation! One tricky thing about "friends" on Facebook is that people do not need to interact with each other to become Facebook friends. Thus, what we can observe on Facebook is that, some people have many Facebook friends who they do not even know and will probably never interact with ever. While for some people, they only add their true friends on Facebook. Also, some people have an account registered but seldom/never use it, so that they will have a few or even no friends on Facebook, but not necessarily mean they do not have a large network. I am wondering how you deal with these problems. And do you think these effects would affect your research and the result?

[YanjieZhou]

Thank you for your presentation! Given that the common criteria for choosing universities are the reputation and the ranking websites such as USnews, QS, etc., the rnaking based on social network is really interesting, but it seems that it lacks the necessary academic components like professors or publishment. So how do you think about combining this criterion with traditional ones?

[chiayunc]

Thank you for sharing your work. In the conclusion, you mentioned that social networks of different classes from the same institution are similar, while same-year networks from different institutions are differentiable, and you mentioned that this could be explained by school characteristics. Is there any generative process that could account for this finding?

[ZhouXing19]

Thank you for your presentation!

[lyl010]

Thanks for coming and sharing this interesting work with us! I am interested in the use of ego-graph and the logic behind random sampling and taking the direct(1-hop) neighbors and the edges between them(page 5). I am wondering can we take 2-hop neighbors instead of 1-hop to construct the network? And will the change of choice on neighbors cause any interesting change in the graph's characteristics? I am curious about whether this kind of details will be a real question and do you have found something interesting about the graph neighborhood distance with the characteristics of it?

I really like the idea to train random forest to tell the relative importance of graph's characteristics, and figure 6 is cool!

[Yilun0221]

Thank you for the interesting presentation! I am also interested in your work in Facebook like @bakerwho. I wonder besides social network, what consist of the most important theories in your work to analyze the users' behavior?

[hanjiaxu]

Thank you for presenting this very interesting research! Moving forward, how would you anticipate such network differences between private and public schools would impact their students' future career opportunities or socioeconomic status?

[SiyuanPengMike]

Thanks a lot for your interesting paper. I have a quick question about the data. The relationships between "friends" on Facebook are actually different. While some people only add their friends in real life, some prefer to add as many friends as possible, even though they are not real friends. How would you deal with these differences in your method? Will it influence your result?

[bjcliang-uchi]

Thank you for this very helpful and inspiring research! I am wondering if you have ever tried other network similarity measurements, especially a deep walk based approach, because it seems that ego-graphs are more about degree centrality and cannot fully capture node's betweenness centrality features.

[MegicLF]

Thank you for this interesting presentation! I wonder if you can discuss more special cases you discovered when you constructed the model and what you did to eliminate their impacts. For instance, a student who does not use Facebook but still has an account may have only several FB friends.

[yutianlai]

Thanks for coming! I'm wondering if you could address more about how network analysis application in industry is different from academia.

[HaowenShang]

Thanks for your presentation! You mentioned that 'Students attending similar kinds of schools have consistently similar social networks on Facebook .' I am wondering beyond the way to represent structural similarities between networks you mentioned in the paper, is there any other way to represent structural similarities between networks? Why you choose this way instead of other ways?

[PAHADRIANUS]

Thank you for your presentation and the associated paper. The dataset created for the project is definitely impressive both in the width of the construction and the depth of the analysis. It set up a great platform for further utilization of the social media on sociological research. Still, though a bit of cliche, it is concerning that how we can avoid walking into zones of infringing privacy when conducting such research. Of course there is no doubt that all information here are already available, but it is possible that conducting analysis may reveal unwanted private info. Additionally, I am curious on how representative is the social media dataset on the entire college students population. After all, whereas people not using facebook might be extinct, among users people have different levels of activity. How to discern such differences and adjust accordingly?

[RuoyunTan]

Could you further explain why you believe "the formation of new Facebook ties closely follows the creation of offline social ties"? Even if students indeed add more friends when they started college and when each academic year begins, this could still be a result of the "People you may know" recommendations if Facebook automatically recommends more new people to users during these times. My guess is that the start of each academic year is when Facebook sees who just added that particular college in their profile. So could there still be any algorithmic confounding issues here?

[jtschoi]

Thank you in advance for your presentation. My question is, would your analysis drastically change if individuals have (or do not have) intent to create social ties using Facebook?

[yongfeilu]

Thank you for your excellent presentation! My question is as follows: When you use the facebook data and publicize your paper, how can you protect the privacy of each individual in your sample? What is the measure you take to protect the private information from being inferred indirectly by others?

[SixueLiu96]

Thanks for this excellent research! I do have a question about the future development of Facebook network cohort. I wonder what would be the future development of Facebook network research? Since people gradually get used to sustain their network cohort by Facebook, what else can we provide to our clients besides this? Thanks!

[harryx113]

Thank you! What kind of impact would the distinction between public and private schools have on building a social network app like Facebook?

[tianyueniu]

Thank you for presenting the paper! I would be really interested to see how would race or nationality play a role in college networks. How do you think they would change the results?

[sanittawan]

Thank you for sharing such an interesting paper. I am aware that this is a descriptive study, but I am wondering if you can speculate about the mechanism that may explain why students at historically black colleges and universities "significantly make more connections" within their respective schools. Would you be able to collect data on race too?

[fulinguo]

Thanks for your presentation! My question is whether the social networks of college students in Facebook are similar to or different from the networks of other types of people, like workers, and whether they are similar to or different from the networks in other social media. You mentioned that your findings are specific to Facebook only, so do the findings in your paper depend on the characteristics of Facebook? Thanks!

[chun-hu]

Thank you for the interesting work! As others have mentioned, people have been increasingly using alternative social apps such as Instagram, Whatsapp, and etc... How would that influence the network formation we observe?

[policyglot]

Trust through Homophily vs Innovation through Brokerage I happen to be taking a course at Booth that's taught by another computational sociologist from Stanford who loves working with network theory (Patrick Bergemann- are you connected?) We're currently discussing how some groups formed during college can promote a sense of belonging without becoming echo-chambers. The Wednesday Ten at Harvard are a prime example: https://www.wsj.com/articles/SB10001424052748704779704574555862616828726

In your data, what examples do you have of strong connections that were formed at colleges (homophily), but where the students went on to work in different sectors and industries (brokerage)? Are there means of the Facebook system of tracking the strength of these cohort networks over time (which allow for the sharing of resources and ideas), other than the fact that there is an edge connecting two people on a social graph?

[timqzhang]

Thank you for your interesting paper ! I want to focus on the external validity of the conclusion of this paper, as Facebook is just one prominent social media platform among various ones. Could it be applied to explain the network in other social media?

[keertanavc]

Thank you for presenting! I noticed that most of your analyses are descriptive in nature performed on a non-time series data (as mentioned in the limitations section). It would be great if you could talk in more detail about how you can extend this study.

[luyingjiang]

Thank you for your presentation. How do you consider a large group of people who do not really use Facebook a lot? Then the network displayed would be incomplete. How these people may affect your research?

[jsgenan]

Thank you for bringing the wonderful work! Everyone must envy the fantastic data you have. I was wondering about the difference between networks formed in the first year of college and those connections formed afterwards. Do they have different structural features?

[Yiqing-Zh]

Thank you for your presentation in advance! It is an interesting topic to study university student networks using FB data. I am wondering whether the algorithm of the FB platform such as suggesting friends distorts your result.