 I'm glad to have Carrie Carr-Howell Youth today, who is a Berkman fellow, and she comes to us from the University of Illinois, where she leads a social spaces group. She's also an MIT alum and studied with another Berkman fellow, Judith and Donna. And we're glad to have her today to talk about text and type strength. So, Carrie. Okay. Well, thank you. I'm just going to warn you guys a bit ahead of time. I probably will go over my 15 minutes, but I'd like to keep this pretty informal. So, if you have any questions, please don't hesitate to jump right in. I'm going to talk today primarily about text and type strength. I tend to show this slide a lot in my talks, mostly because I think we forget just how important people are, and we're all here, and a lot of our technology actually deals with how people interact with each other. And the bulk of my talk is going to be about the connections between people. This is a quote by William White, who is a sociologist who actually studied public spaces and how people use public spaces. One of the interesting things he noted was that people flock to spaces with light, with trees, with water, but the number one thing people wanted was a place to sit. And the reason I bring this up is how often we design spaces is if we don't want people to actually hang out there. What is this equivalent of the seat in a virtual space? So, bulk of my talk is going to deal with text. This is the first known text that people have found. This is actually a Sumerian text from 3300 BC. And what's interesting about this is it's archived, not maybe archived the way things are archived online, but it's an archival version that we can still peruse today and we can still make inferences about it. A lot of the early work that I did when I was a graduate school was some of the early work on sentiment analysis. This actually is a snapshot of the Usenet group, social.culture.grease. And what this was showing actually on the right is angry messages were red, peaceful messages were green, news was another color. So this was a very irate group and again, this was all influenced by text. A lot of the later work that we did, I'm not gonna talk about today, I'm just gonna talk about ever so briefly just so you see what kind of work that my research group does. Everything we do deals with communication. In this particular case, we're doing visualizations of people on a tabletop. Each person is a different color and what we found here was that if you show people visualization of their contribution, they tend to balance interaction. We looked at turn taking and we're actually using these visualizations to teach turn taking to children with Asperger's. On the lower right, you can see the snapshot of somebody talking and then just one little sliver of color in there. That's somebody going yeah or aha. It turns out people make stories about these visualizations. So for example, they say the red person is the leader, the green person is like the follower. This visualization here on the left, this might be indicative of me speaking today. I'll be blue and I'll be totally dominating this conversation, not letting many other people get a word and most probably done other visualizations where we incorporate voting into the visualization. One of the things that we found in these is that even if outcome people want doesn't actually happen, they like the idea of having their vote actually embedded into the visualization in real time. Almost as if like a carry was here feature. I was here, I was listened to, I'm present in this community. We did some other visualizations where we actually did archivals of meetings and this actually is a textual archive where you can see one third of conversations are talking about downtown, then lunch, then social. Another part of the conversation drifted off, split off, talked about wine and we actually had to build our own topic clustering for this because it turns out lots of existing topic clustering today is built off of the Wall Street Journal. And people don't talk in third person, they don't always talk about the Iraq war. But talking about Thai strength, a lot of this work was influenced by our early work looking at how different people in different areas use social media. So for example, I grew up in a very, very small village in Greece. At the time it was about 1,000 people, right now it's probably 300 people. And one of the things that struck me when I was there was that every Sunday my dad who was in the US would call it noon and I would wait at the one phone call in the village to get that call. It was in a tavern so they'd sort of like liquor there, coffee, people would be paying back, amen. So every Sunday at noon I'd go wait for that phone call. And about 10 to 15 other people would show up, listen to my half of the conversation. It was almost as if that was advertised entertainment there. TV didn't start until 5 o'clock. But this one phone was kind of like this hub. Everyone went there and if you were bored you just kind of hung out there to see who might call. Later on in 84, my grandfather got the first phone in the village and it didn't make that much sense because there was nobody else to call. So what happened was that people started showing up there waiting for phone calls from the US. So the house became a nice hub. But one of the things that I want to stress with this is how people use phones differently in different parts of the world, especially rural and urban areas. In the United States rural and urban areas use the phone very differently. In fact, most people don't realize that prior to 1920 there were more phones in rural America than there were in urban America. And in looking at the telegraph, looking at the telephone, looking at the internet, people have to use them differently in rural and urban areas. And one of the things that we wanted to find out was why and how. So just going slightly to definition of rural, rural town is a town with less than 2,500 people not connected to a metropolitan area. Urbana is urban by this definition. We're quite close to Chicago. We're about 40,000 people. Champagne's about 70,000 people. Just to give you a feel for what this like around the world, 24% of the US population is rural. 50% of the world's population is rural. And this image here isn't rural versus urban. We're just to give you a context of looking at the lights at night, just how much of what little part of the world is actually lit up at night. So one of the things that we were interested in doing in this work, and this work is actually done with my graduate student, Eric Gilbert, wanted to see how people use social media differently in urban and rural spaces, just to get a starting off point to see what these interfaces should look like and compare them to what they do look like. So we decided to just sample random zip codes from rural and urban areas. This is an example of one of the rural areas that we looked at. And this is 100 West Virginia, population of 344. And we wanted to see just what parameters people use differently. This is an example from, I love this picture. It's from Claude Fisher's book, America Calling, Social History of the Telephone, America Calling, Social History of the Telephone. And it shows pictures of people using the phone and how the phone actually makes farm life enjoyable. Not only that, but the old-time isolation and loneliness of farm life is a thing of the past once you have the telephone. At the time, actually, rural people loved party lines. This idea of, like, it's almost like a subnet, if you think of it in terms of networking terms. You could pick up the phone and almost listen in on what other people were saying. Urban people decided that this was passe, this was not the way telephones should be used. And telephone companies agreed with them in many ways. This is a quote that kind of illustrates this, where it says, farmers as a class are troublesome customers to handle, and are apt to have an exaggerated idea of their own rights. The bumpiousness of certain farmers can be overcome only by constant efforts to educate them. This is by TelecomExec in 1918, also from the Claude Fisher book. So what we did is we sampled MySpace, in particular, 3,000 public users, total of 340,000 online friendships, 200,000 interpersonal messages. This is a sampling, just so you get a feel for the zip codes of the US. This is a sampling of what we did for our study. The demographics that we had were about roughly even rural to urban. One of the things that I want to stress, we look at the MySpace ID. The reason that's longer is because the later you log into MySpace, the larger your ID number is. You can see the difference between how often rural people, actually, when they logged in versus when urban people logged in. Ages were roughly similar. Days since last logged in, quite different. Rural people about four days. Urban people about 10 days. So we had about five hypotheses in doing this work. And these hypotheses were actually built off of literature and theories from face-to-face work in sociology. So our first hypothesis was that people in rural areas will have fewer friends and comments on MySpace. Second hypothesis was that there will be more women on these social sites. Third hypothesis would be that there was more private profiles. And the fourth hypothesis was that the friends would be closer geographically. Fifth hypothesis would be a preference for strong ties over weak ties in rural areas versus urban areas. And one thing just, I mentioned the telegraph earlier. Every time this new technology came out, it was supposed to change the world, and it was supposed to democratize interaction conversation. When the telegraph came out, it said people said war would end because people could communicate and there would be no misunderstandings. When the airplane came out, people felt that you'd have no more war because people could just drop a bomb and who would ever do that. But a lot of this came out with the internet and with social media as well. So I'm not going to go into the depths of this study, but I want to tell you what hypotheses we proved and what we couldn't prove to lead into our tie strength work. So for example, we did find that, yes, rural people did have fewer friends and comments. Rural people had more women involved, and not only were there more women involved, but women were also the keepers of privacy in these sites. There were more private profiles in rural areas, in rural profiles, and their friends were geographically closer. What we could not prove or disprove was that rural people had a preference for strong ties over weak ties, as is the case in face-to-face interaction. There were several reasons we couldn't prove this. One of them is that there isn't a really accepted quantified approach for what is a strong tie and what is a weak tie. People use us quite a bit in literature, but it's very hard to use in practice, and it's hard to actually generalize what it means to be a strong tie versus a weak tie. One of the things that we found from this work was that the rural and urban areas were very much disconnected in terms of where their friends were. It's kind of hard to see in this picture. But one of things, and this may sound kind of extreme, but maybe there are two mice spaces, you know, a rural mice space and an urban mice space. And what can we do as designers and technologists to actually bridge that gap? Because we found that people in rural spaces wanted to communicate more with people online. They wanted to have more conversations. They wanted to friend more people. But there were issues of trust. But going back to that issue of tie strength, we decided, because we couldn't find a good metric to actually come up with to prove or disprove our hypothesis, we wanted to interrogate that a little bit more and try to figure out what tie strength really means online, how different tie strength is online from offline. So we decided to explore Facebook. This is a snapshot of my Facebook profile from way back. And one of the interesting things about Facebook is that you see a list of people. But people are essentially friends or strangers. They're on that list, or they're not on that list. So it's quite binary. And someone who I talked to, who's on this list that was a dog breeder that I've maybe spoken to once in my life, has the same presentation as my husband, who I hopefully talk to every day or more. Just as a snapshot of what my Facebook profile looked like this summer, this is what the connections look like. And this is made using a structural tool called Guess. And I kind of like this image because it looks like a question mark. One of the things that's interesting about this is that there's like three different people down here. These are my only three relatives in Greece with the computer, just also to show you the difference between a rural and urban space in terms of using social media. And everyone else, I mean, they're disconnected from everyone else. And one of the things that I'm showing this image to stress is that a lot of this is actually built on structural data from Facebook. And in terms of full disclosure, I should say that just having this image up here actually does violate Facebook's terms of service, which is a different issue I'm not going to get into today. But in terms of the ethical implications of what this means. But tie strength. Tie strength was probably coined by Mark Renovetter. I would say in the 70s. He'd been studying this way before that. But his definition is the strength. What his claims about tie strength are that the strength of a tie is probably linear combination of the amount of time, emotional intensity, and intimacy, and the reciprocal services which characterize the tie. What does this mean in layman terms? Strong ties are people you trust. For example, if I wanted to borrow $100,000 from someone, I would probably go to a strong tie. My parents, my brother, my husband. Also, strong ties have been shown to actually help people in corporations get through difficult times. Strong things help people in grieving situations if they've lost somebody. Strong ties also can have the opposite effect sometimes. If somebody in your social network is a strong tie gets depressed, there's a higher probability that you might get depressed as well. Weak ties, on the other hand, are acquaintances. So for example, a friend of a friend, someone you went to high school with, a college buddy. And for the longest time, people thought that to lead a healthy, protective life, all you really needed was lots and lots of strong ties. And then Mark Renovator wrote this really great book called Getting a Job. And what he wrote in his book is how he emphasized how you really need weak ties to actually make this happen. Because your strong ties now all the same people, whereas your weak ties, the tentacles of them get dispersed more throughout the continent, more throughout the world. And you actually get more information dispersed that way and new information that you've not have gotten from your strong ties. So this work, I want to say, is also done with my PhD student, Eric Gilbert, who's on the job market right now. So what we did with Facebook was we wanted to look at what tie strength meant. So we looked just as an aside, there's over 7,000 papers at site, the strength of weak ties, the seminal paper, might mark Renovator. And at the end of this paper, he hints at this continuum between strong and weak. But doesn't so much explicitly state it. So a lot of the work to date actually says, this is a strong tie, this is a weak tie. And they come to these calculations by saying how many times people have spoken to each other. And if they've spoken to each other over 10 times, that's a strong tie. And if they've spoken to each other less than 10 times, that's a weak tie. We wanted to see if we could tease out a bit more and maybe find out if this is a continuum and if so, how it's a continuum. So adding onto the existing work, and Renovator discussed in the previous quote, intensity, intimacy, duration, and services, we really wanted to add a bit more and try to add onto the existing work. And we're intrigued by Wellman's ideas of emotional support and his added work on intimacy, the idea of Nandlin social distance and Bert's structural connections. And actually here's a really good new book out right now. So essentially our task was looking at Facebook. How can we take Facebook and actually map something into a face-to-face environment? In a way to do this, it's practical, that's interesting, and actually can influence design decisions in our work. So what we did is we brought people into a lab setting. And I'm not going to get also into the ethical implications of sampling and doing all this online work, but we brought people into our lab and we ran our algorithms and we had them actually choose, fill out these forms that were designed using Grace Monkey. So for example, we asked them questions such as, how strong is your relationship with this person? How would you feel asking this friend to loan you $100 or more? Yes. I'm going to take back one slide. What exactly are you trying to map? What are the two things you're mapping? So we have these, we took over 70 different parameters from Facebook and we're trying to map them to Thai strength, how strong somebody's relationship is. So for example, looking at all these parameters, we're going to come up with a number between zero and one that tells you how strong a Thai you are between this other person. And to help us build our model, we needed something to map our algorithm against. So the third question we asked was, how helpful would this person be if you were looking for a job? How upset would you be if this person unfriended you? And if you left Facebook for another social site, how important would it be to bring this friend along? One of the interesting things that happened with Friendster, if many of you remember, is that it kind of disappeared in a very short amount of time and we're interested in that. Did people bring their close friends with them somewhere else or did they just sort of like just scatter and regroup someplace else? But for the purposes of this talk, I'm going to focus on this first question, which is how strong is your relationship with this person? So what we did is we assessed over 2,000 friendships and this is from 35 university students and staff. So of these 35, 23 were female. On average, there were about 150 friends per participant. And again, we looked at over 70 numerical indicators that we scanned to build our model. I'm not going to go over all 70 of these in detail because you'll get really, really bored if I do that. But I'm going to give you a sampling of the ones we used in seven different categories. The first category is intensity and we looked at there was we looked at initiated wall posts, friend initiated wall posts, wall words exchange. We also looked at the inbox, which is why we brought people into the lab and we didn't do this using a Facebook app. So we looked at inbox messages together, thread depth, status updates from friends. In terms of intimacy variables, we looked at participants, friends, day since last communication, wall intimacy words, index intimacy words. And we did a lot of this using sentiment analysis. We looked at predictive variables of social distance and these are age difference, educational difference, political, occupational difference. The traditional structure stuff that many people do. We looked at reciprocal service usage links exchanged by wall in terms of reciprocity, applications people had in common, positive emotion words, negative emotion words in terms of emotional support. And we looked at duration because duration really does play a key role and that's probably one of the reasons people have been using it as a proxy for so long. And I'm going to skip to show you right now the results of some of our work in terms of what the model told us. So these are the seven different classifications of things that we looked at within Facebook. So you see structure, you see emotional support services, social distance, duration, intensity, intimacy. And although not one of these monopolizes our model or the strength of our model, it was really interesting to me to note that intimacy which is this one right here plays such a big role in it. Alone it probably wouldn't be the best indicator but when you build this together and you look at something like intimacy and something like emotional support and structure which was primarily used in the past, you can just see how much more this idea of intimacy actually adds into the connection between people than without it. So what does this mean? So we came up with these variables. One of the things that we also found was that if people want to build something from our work, using these 70 plus variables doesn't make that much sense especially if you want to generalize. So we found that if you take the first 15 of them, it turns out you get more than 50% of the power of our algorithm. And many of these just you get a feel for one of the most predictive elements where day since last communication, day since first communication, intimacy time structural and if you want to go into more depth in that, yes. Just a point of clarification, you're using the quantitative data from the Facebook profile to determine the qualitative responses of the survey. Exactly. Yeah. By the end of the day, since first communication, should be such a strong predictor of sentiment? In my case, like for example, if you have a very strong tie, like my husband is a strong tie, we don't communicate with each other on Facebook at all. And it literature suggests and shows actually that strong ties connected, I've talked to you on a various platform. So we talked to each other face to face, we use the phone, we use email, but day since first communication probably was a big indicator for us. So maybe when you first log into Facebook, the first people you log in, it's the first people you connect to tend to be your very, very strong ties. And what's interesting is that even though he and I communicate almost zero on Facebook, you know, it could still bring that out as a connection between us. You mentioned intimacy in its own, it's not a very good indicator. Can you mention a couple of words why? Well, that's an interesting thing about building models, is that when you have different parameters, you know, just by itself, it's gonna give you a good proxy, but it's not gonna give you the most accurate story. So when people were using, you know, duration of conversation, like how much you talk to people, like a standard thing people have done, us included, was just saying look, if you talk to somebody more than 10 times, it was a strong tie. If you talk to somebody less than 10 times, that was a weak tie. It's also not a bad proxy, but again, that leaves out intimacy, it leaves out all these other factors. We tried to bring in several of these things together to see if we could get a better indicator. And I should have had the slide, I didn't bring it, but in terms of using the previous techniques, your results are like maybe something like, you know, 50 to 60% of, you know, is this actually a strong tie versus a weak tie? Whereas our results of strong tie to weak tie are 87%. Yes. The intimacy made up 37% of, what does it make up 37% of? I didn't quite understand. Oh, the power of the model. So it turns out that the intimacy features give us that much, add that much more accuracy to our model. Like what are you measuring it against to know that your model is 30% accurate? Oh no, of the total, of the total. So of all the things that make our model work, that accounted for 37% of it. Does that make sense? I'll let it go. Yeah, okay. We can talk about that more in depth if you'd like. So the reason I have this slide at Peter is that if somebody wanted to make a quick and dirty version of this, they don't have to use all 70 plus features. They could use, you know, a subset of them. But one of the things that I found interesting about this was not just when it worked, but when it didn't work. So this is actually a quote from one of our subjects. She said, I, yes, this friend is an old ex. We haven't really spoken to each other in about six years, but we ended up friending each other on Facebook when I first joined. But he's still important to me. We were best friends for seven years before we dated. So I rated it where I did. I was actually thinking of rating it higher because I'm optimistically hoping we'll recover some of our best friendness after a while. Hasn't happened yet though. So this is an interesting case where you have like the dumper versus the dumpy, whereas the person who, you know, did the dumping usually evaluate somebody higher than somebody who actually was hurt by the situation. But this also brings up another interesting point in that tie strength is strong and there's a very, very fine line between love and hate. And hate can also be a very, very strong tie. And one of the reasons that I mentioned this is because we really need to start looking at the strength of positive and negative ties. And Kleinberg is starting to do some really good work in this area. But maybe our algorithm didn't tease out the positive and the negative at very well. Maybe it brought, we have this assumption here that it's a positive strong tie and that's what's built into our model. Another example of where this didn't work is when people have these weird feuds and they use people's profiles as proxy. So for example, we were neighbors for a few years. I babysat our child multiple times. She comes over for parties. I'm pissed off at her right now, but it's still 0.8. Her little son, now three, has an account on Facebook. We usually communicate with each other on Facebook via her son's account. This is our one mutual friend. And this is fascinating to me for several reasons. Like one, because technically you need to be 13 years old to have an account on Facebook. And two, in many ways, are using this as their proxy. I mean they are communicating. Our model will not pick this up because it assumes that, you know, it assumes that when you're honest, you're saying who you say you are, you're communicating with people. You're not using somebody else's account to communicate. So again, I'm not surprised it didn't pick this out. But in many ways, I think where our model breaks is far more interesting than where it works. So why do we care? Let's say we have a model for predicting tie strength. And let's say I can tell that, you know, my tie strength with you is 0.5, whereas my tie strength with you might be 0.7. You know, where does this matter? How can we use this? So one example of where this might matter is in terms of, you know, possible friending scenarios on Facebook where people are, you know, maybe you should friend this person, maybe you should friend that person. The typical standard thing that people have done in the past, we wanted to see what it means in the world of design and the world of applications. So for example, in our earlier work when I talked about the rural and urban spaces, you know, how can we design differently to account for different types of users? Maybe we don't need one interface for all. Maybe different people need different interfaces. How can we use this work to influence design in Facebook? So one example might be in the privacy settings for Facebook. You know, I'm a relatively intelligent person. You know, I have a doctorate. And getting through all these privacy settings in Facebook can be somewhat of a nightmare sometimes. What if you were to use something like this to actually start organizing your photographs? So for example, you know, your strong ties saw a different subset of photographs. Your weak ties saw a different subset. And then you can drag and drop in between or make even more categories that you might want. And again, this is just to make your life a little bit easier. In our average, you know, people had roughly 150 friends. In my students, for example, have many, many more friends in that trying to make a profile for 500 people to see individual photographs can be quite, can be a nightmare. But not only that, okay, so we built this model and this model happened to work for Facebook in 2008. What does this mean in 2010? What does this mean in 2020? And what does it mean in other online environments? So we wanted to see, you know, can this generalize to any other different world? And we were infatuated with Twitter at the time. So we decided to see how can we use our model? Will our model work with Twitter? So this is a different social site. I know many of you are on it because I follow many of you. So Twitter is quite different from Facebook for several different reasons. I mean, one is a very different model. It does not have the reciprocity that Facebook has. But it has a lower barrier to entry in many ways and it has a huge population. And so it seemed like an interesting approach to see, well, let's take our model, adapt it to Twitter and see how well it works. So what we did is we took the same parameters that work in our Facebook model. So we basically built a model on Facebook. We took those same things. We morphed them onto Twitter because not everything maps perfectly. For example, you don't have an inbox in Twitter to look at intimacy words there. And then what we did is we built an interface. So I'm gonna switch into live demo mode now, which usually is the kiss of death for me, but I'm gonna try it anyway. But I'm gonna do that because one, it's gonna be far more interesting for people to actually see what it does instead of me just talking over a slide. So for example, this is my WeMetal account right here. And I don't have that many friends on WeMetal. I would classify myself as on Twitter. I could classify myself as a lurker. I'm addicted to it. I adore it, but I tend to be more private than many of the people that I follow. So I'm gonna try to reload this again, sign in, and it verifies actually your login with your Twitter account. Okay, so what do we see here? So I've logged into WeMetal and this is the, we built our own different list application and clients. So I'm gonna show you two different things here. The first thing I'm gonna show you is the list and then I'm gonna show you is the client. But what does this list thing do? So Twitter recently announced Twitter lists and they make basically lists of people you can follow. You know, Twitter was getting kind of cumbersome. You could follow maybe like thousands of people and how do you get through all of these tweets at one time by having these lists that you can create or you can even follow other people's lists. You can actually have these interesting types of filters that you could look through. So what our algorithm did is it made you a list for an inner circle and the inner circle are my strong ties. So for example, Eric Gilbert, PhD student, he's here, a colleague at UIUC is here. This is Christian Sandvik, fellow-fellow. This is Lisa Nakamura. I'm not sure why this showed up as a strong tie for Kai and this is Joni Dumeco at IBM. These are people I follow. These are people. Should you make these for lists or did we middle do it for you? We middle did it for us. So we middle created these lists. Down here is my outer circle. So people that I follow that I'm not very connected to. So there's somebody here. We'll work on that. We'll work on that. I mean, I am ready to talk about the moment, so. What we have here on the right is it builds communities for you. So it uses, in my case, because I don't follow that many people, it only built two communities. And this first one here, and you can change these names. So for example, in this one, it's probably, in my case, a research community. We made up that name for you? That's a default name we give people and then we let them actually change it to make it what they want. And so if I call this research, when I then log into Twitter, there'll be a research list for me that I made based on these people. So there are some familiar faces here, like Esther is here, Dana is here. This is a different community. So again, my student is here again and we tend to publish in the same conferences. If you go here, you start seeing easy to miss people that don't tweet as much, but maybe there are tweets that I wanna see because they don't tweet as much as the norm. And here are eager tweeters, people who tweet a lot. And I have to say, I'm in love with Roger Ebert at the moment. I think he's just one of the wittiest Twitterers out there. But this is what my profile looks like. And my profile is a bit meager. We decided to add to this our own Twitter client and please be patient with us. This is the second version of it out. It's been out now for maybe about two weeks. This is the second incantation of it. We had one out in the fall. There were about 600 users. And we wanted to see, explore how people can actually use tie strength to look at their tweets. So for example, oh, here he is. This is Ebert. So I don't have that many strong ties up here. The weaker the tie, it goes into black and white. If it's a very strong tie, then the picture also appears bigger. So if I move up here, this is what we call, we try to use a different type of term for zooming. We call this social zooming. There's also different types of zooming. And so let's see if this will work. I'm sorry. Increase text size. Oh, increase text size. Let's see if this will, my mouse back. Groups are formed by doing some clustering on those parameters, the variables. We have this spectrum of strong and weak tie and in this particular case. So what this means here in a sense is this is my inner circle and this is my outer circle. So if I move closer to my inner circle, I get my strong ties. If I move closer to the outer circle, I get my weak ties. This here is a time mode and this is a collapsed mode. Sometimes between, because like the eager tweeters tweet so much, you don't want to open your page and see 20 tweets from the same person. You'll see like a highlighted tweet and it'll say like plus 10 more. Let me see if I can increase this. I never increased the size of my... But those different groups that came before, how do they relate to this bar? In this case, they will relate when you go to regular Twitter. But the groups that were in my inner circle will be the ones that show up up higher and the ones that were in my outer circle will show up lower. Thank you. Is that better? Okay, great. So in terms of my strong ties, if you've noticed from my list before, these are my only two strong ties that have tweeted. And if I go to my weak ties, oh, okay, Carrie's talk is beginning. So my students are watching this talk. If you go here to the weak ties, that's what's being said there. The interesting thing for me in playing with this interface for these last few weeks is that I'm somewhat more interested in what my weak ties have to say than what my strong ties have to say because we're in the same community. I pretty much know what they're gonna say. Not all of it, not all of it. But the idea here is to be able to actually explore and move back and forth between the different ties and see where people set it. So right now we'd like to see where people like to keep this bar. We also added some other features. For example, because it builds communities, you can show only your first community. In this case, it would probably be what I titled research. And there's a good chance this is crashing right now. The community is the most the same thing. I'm sorry? The community is the most the same thing. No, they're not list. A community is actually a group of people that were decided to be a cluster using sort of Markov clustering in our case. But we made a community a list. So it's not the same thing, one of the lists we defined was a community. So in theory, what this should show if it worked was positive tweets, meaning things that were positive emotion and negative emotion. People that post frequently, oh, there's a frequent poster. So that actually worked. And then people that post infrequently. And what do you use for sentiment analysis? We use primarily the work done by Cardi at Cornell. So it's a pretty simple approach and it works really well and really fast on Twitter. So I highly recommend it if you want to start playing with sentiment. We've had some complaints about how things that are negative aren't really negative. But before I move on, actually, let me just log into Twitter really quickly so you can see that the one list I made actually shows up there. So this is my account. So the research that I just renamed actually shows up here. And you can publish it, the list, the inner circle actually showed up here. I like to follow Nancy Bain's list. So basically, she made this list and I just decided I was going to follow it. So it actually integrates whatever you make into your Twitter feed. But just to show you something maybe a little bit more interesting, my student who follows many, many more people than I do, it makes it for a much more interesting. And this is Eric Gilbert's account. So you can see that his inner circle is much more populated. And people like to explore and try to figure out what this group is right here. So my screen is a little too big, so let me try to move this. So for example, trying to figure out what makes this a community versus what makes this a community. So some interesting things that we've encountered so far, just from asking people and from people leaving us suggestions. Sometimes if somebody shows up in your inner circle, that tends to be an X, people get very, very upset. So this is a recurring theme again. Like, why is this person showing up in my inner circle? Well, why is the problem? Maybe, maybe. So you can delete people. I don't want to delete people from Eric's inner circle right now, but you can see again who these people are. If you click, you'll drop this person from a list. One thing we'd like to do is add a feature to add people to a list instead of just dropping people, because they actually make it much more useful. Drag them down by the outer circle? Not yet. That's one of our goals. That's one of our goals. And again, if you look here, because he has such a bigger, he follows so many more people, he has a bigger community. So he has an extra community, another birds of a feather community. And what's interesting about this one is that most of these people had had something to do with the UIUC. So you can start guessing what some of these communities are. And you might rename them and say that that's the UIUC community. And again, in his case, the easy to miss people are here and the eager tweeters are here. And so one thing I'd like to do before closing this talk is I just ask one of you to log in. Maybe Ethan. I have logged in. You have logged in? Really? OK. No, it's sort of astounding and scary how good it is. So in your experience, the people that it shows for your inner circle, how well do they, how well do you think they fit into your inner circle? I would say the inner circle is about 70% accurate, maybe 80%. And it does actually a fairly, well, it's such an interesting question, right? So it includes my wife, which is good. And people like Ivan Siegel, who runs Global Voices, which is good. My baby blogs but doesn't tweet. But mostly what it is is the inner circle is sort of a blend of my immediate, personal universe and probably my two most important social universes, which is to say the sort of global voices space and one chunk of the Birkeland space. What's actually most interesting is that it created two groups. And the two groups actually very neatly become sort of cosmopolitan, right? The sort of the global voices, the coolest Bahraini blogger out there, the coolest Malagasy blogger out there. And then another one, which is very clearly the straightforward all-American white guys to Gerati. And it just sorts the two very, very straightforward with almost no overlap between the two. It's actually, it's very, very impressive. Cool. Do you think the list suggesting people you should be following? No, no. It's somewhere on our 100 list of things to do. But we haven't done that just yet. So one of the things is that we want to study how well this generalizes. One way to do that is see if people remove people from their inner circle and outer circle and who they add. So far, I mean, because you can only remove people, people have removed less than 1% of people on their inner circle or outer circle. But our goal with this work actually is to do a quantitative analysis by looking at the logs, but also do a qualitative analysis and interview people. Because when you interview people, you actually get at some of the gist of what's really going on that you don't get at just by looking at server logs. So that's one of our next steps. Also, we want to see how people use the client. Do they like the client? One interesting thing about developing a client when there's just so few people working on it is that you can't compete with something like TweetDeck, which is a whole company. So we put a lot of effort and design into the fonts, into the layout of the client. You can also tweet from it. I didn't make it clear before. It's kind of, let me go back and just. So when you go to the client, so there's a little tweet button up here. You can actually use this to send a tweet. And it will say on Twitter that it was sent by WeMetal. One thing that I would really love one of my pet peeves from all the four tweets that I've sent is that when you send a tweet, you want to send something, you have to go all the way to the top. I would really love it if I could, as I'm reading something, just send a tweet right then and there in line as opposed to having to move all the way up here and on Twitter and on our interface and actually type the tweet up here. But basically, I'd like to open the floor for suggestions because we're very open to criticism. We want to make this as best as we can before we start studying it. And just ideas that people might have for how it might work in applications that you're working on or in research agendas that you're interested in. MySpace is that, so is it taking a linear combination of those? So what we did, we didn't do anything on MySpace. We did a model on Facebook and that's, we did that using, oh, linear regression is what it is. And have you tried other types of models to see how they compare? Then you're using the same model here, WeMetal, same model. Checked how well it tests it against. Yeah, that's our plan to find out in the real world with many, many people. So we have about 600 users now. We hope to get up to a few thousand and then see what we get by having lots of numbers of people use it. In terms of the visual feedback, I'm not sure if that's a priority, but as far as I'm concerned, geolocation and perhaps the ability to navigate and see geographically who is where and then how from account to account, but as if you were seeing the world from on high. So if I could see you, you would be in Chicago and then like, you know. So you mean like some, a map like visualization? Yes. For Twitter? For Twitter, or for a manual? Someone just did a really nice piece of Facebook visualization around this, looking at essentially clustering people's Facebook friends based on looking at their public profiles and then trying to figure out geographic clusters and basically figured out that San Francisco and LA are very tightly tied. Basically Seattle and Portland know whatever escapes that orbit. There are areas where there are a lot more cross-national friendships and then there are areas like in the Northeast where we basically don't get out of our orc and states or counties. I would love to see that. I've never seen that visualization. I'll send it to you on Twitter and then our tie will get stronger. Thank you. Can you send it to Berkman Friends? Yes, I will. I'll send it to you. It's gonna go out to the Berkman Friends list and if I could make it appear over my head, I would find a way to do that. So David, when you're in your list out of curiosity, was your inner circle, did it make sense? So yes and no. Okay. So there are, it's trouble, it's hard for me to figure out who's the closest tie and who's the weak tie. So, and why some of the people are. So, I follow like 30 and 20 people. So that's more than I actually follow. And that may be part of the problem. There aren't people in it that I want to say no. I don't, I hardly know that person. So we did a very good job there. But there are people in the other groups that I say, well, I would have thought they would have made it into the circle. So part of my reaction to this, I mean, I really like it. I have a very quick question about it. The part of my reaction to it is I sort of want to know more about how, just as a user, what does inner circle mean? So, if there are a way for me to click on this and to see what you're doing, I'd be... Click on. A link on the page that tells me. Oh, information. Information, yes. That's what I'm talking about. Information, yes. And the very quick question, is this tweetable or logable? Yeah. Okay. Yes. Yeah, I also did it and find it a little puzzling that I came up with. I personally would be really interested in trying it on Facebook. Yeah. And at least in my experience, to me, Facebook and Twitter are really, really different in this notion of the inner circle, which makes lots of sense in terms of Facebook. I'm not sure. I think there are people who use Twitter in a Facebook-like way. And I think, particularly if you're doing it in a university environment, a lot of students do because they're following each other around the campus and it is a little bit more social networking. And I think what you're finding here is this is a community that uses where Twitter is more about broadcasting news and getting news broadcasts and that notion of inner circle is very, very, it doesn't matter as well. And so I think in some ways, I love the idea, but I think it may have been a little too untranslated in its movement from Facebook. Well, there's another point there, which is that with this version, no, we came up with one generalized model in Facebook and we decided to take that model, transplant it into Twitter. But as we're getting feedback from lots of different people about what they move around, we could make individual models for a person or we could make, again, another generalized model for everybody. And what do you do? And is tie strength one general model? Or are people's experiences and how they use this system? For example, I'm a low-traffic user. Should my model be different than maybe Ethan's? Right, but I think, for instance, and some of this may be data that you can't get because this you can't get off of Twitter, but one of the most, especially people who are using it as a portal to news, is not even how much you retweet something, but how many times you follow the link that someone has set off. So I think if you're setting up your own client, if you can start collecting things like, if I tend to click on links that this person put up, if I'm using this as a news service, that's a huge piece because the people who sent out links and I don't follow them, it's not interesting. So I think thinking about what makes, it might be interesting to do it as a paper too, what makes Twitter different as opposed to how to make them look the same. Yeah, yeah. You know, the projection you did into kind of a strong and weak work pretty well for Facebook? I'm wondering if there are like multiple things going on on Twitter. And so the projection into two-spaces doesn't seem quite the same. Yeah, yeah. And like I said earlier, the spaces are so different. I mean, we took advantage of reciprocity in Facebook and that doesn't exist to the same extent in Twitter as it does in Facebook. Now, one of our goals is to see just how much it generalizes. Like if it gets things accurate to some degree, we can say that, look, with these parameters, you can get a good gist of what tie strength is. And that in itself is an interesting finding. Finding the ideal tie strength for Twitter is a much different story, but... Just in terms of thinking of spaces for the research, well, first of all, I'll undo this comment. I think this whole question of sort of inner circles, I think all of this gets really, really mucky. Yeah. And it's really impossible for an inner circle on something like this to sort of reflect my emotional inner circle, because perhaps, thank God, most of the most emotionally relevant people in my life don't tweet. So I find myself looking at it instead of going, are these the people whose tweets I take most seriously? Is this a list of people who I generally, who I care about, independent of the content of it? It's a nice provocation. I do think that there's gonna be some sort of vagueness that comes into play no matter what you do. I just wanted to suggest, as a research direction on this, particularly based on the Facebook work, as you completely correctly pointed out, Facebook gives you no valence. It's friend or not friend. Live Journal gives you beautiful abilities to do valence. And I'm married to an extremely heavy live journal, Dream With User. And watching her sophistication in carving up communities and then publishing information selectively to those communities is just fascinating. And so I guess part of what I wonder in all of this is that you've taken two tools that have sort of very heavy artifact behavior too. Like Facebook has the artifact that every friendship is reciprocal, mutual, as strong as it could be, et cetera, et cetera. Even if you don't have to reciprocate, you do in Facebook. Twitter has moved into the sort of celebrity position where once you're over a couple thousand followers, you're probably not gonna detail your most intimate thoughts. You're using it as a broadcast platform. It might be really fun as sort of a third avenue in this to look for something that has taken precisely this issue of strong tie versus weak tie group formation and made it essentially central to the technology instead of see how it plays out. Yeah, that's an excellent point. You know, the murkiness I think is one of the reasons why people have just kept it in two bins for so long. Just strong and weak tie. And in some ways it's almost an art form trying to figure out what is strong and what is weak. And there hasn't been an algorithm for, you know, one thing that I should stress is that although we decided that we're gonna talk about a continuum, we still put it into two bins. So I would like to further explore what it means to be in the middle. I mean, just as a for instance, if you looked at a network like LiveJournal, what you're gonna find out is that there are sub-communities where my wife basically maintains a public list, a close friends list, and then a list for a specific sub-community that she's involved with. Any given post on any given day could end up in any of those three. But within each of those, there are stronger and weaker ties. And it would be interesting to think about whether you can pull it apart. For me, one of the most interesting things I think I'm thinking of looking at right now is that if you look at a lot of these small world problems, like Nellyn did a lot of studies. Like hers were different from Milgram's and that they were, his were different from Milgram's and that they were in one city. But looking at, he was looking at the tie strings of the people that got the packages. So I may get this background, but I believe that in the beginning they went out to weak ties and by the end they went to strong ties. So trying to see, there's so much literature on information flow, but I don't think there's been as much looking at, what do you send to strong ties and what do you send to weak ties and why? And how tie strength as a variable influences information flow across a network. Like I think that's another area that might. I was gonna get back to the point that you were making about Facebook being more social, Twitter being more informational, promote him, necessarily everybody, but given that possible dichotomy that one way to analyze the Twitter differently might be on the, not so much the strength of the tie, but the rated value of the information. And so that, you know, and then possibly well, given what they say about weak ties, that you get more different information from the weak ties, maybe you actually find something there about, oh a lot of the most highly rated information are from some of the weakest ties, but they'll come up with a different. Yeah, like actually I was thinking like something like Reddit or Dig would be a good way to look at that because you can actually people vote on like how much they liked something. I mean just getting back to the definition of the specificity of the Twitter model is that another part of it when you think about saying, well we wanna move beyond strong and weak ties is also to think about what types of relationships exist in these online ecologies that just there isn't a model for very well in our day-to-day life. And Twitter is really interesting that way because for instance there's all these people who I might follow them but they don't follow me and vice versa. There's like when I, then the part of the, I think it makes it an interesting interface problem is because even though I know that intellectually when I post something I keep thinking that I'm posting to a group of people who I follow but the group of people who are following me is a very very different group and so it's not the same one and it's a very bizarre set of ties and I think thinking about how you can use this type of thing to help clarify a type of relationship that we don't have a good model for understanding. Can I pose a question to the crowd? Like how do you all reach Twitter? Do you just go to the very top and start reading? Do you use TweetDeck and do you use lists in TweetDeck? Like how do you handle? I'm guessing a lot of people here follow a lot more people than I do. So I wonder, I am curious how many people have this same experience. I do where I tweet but honestly the amount that I tweet isn't a lot to begin with but I definitely don't bother to read anywhere in here. Really? Okay, so you produce more than you consume? Like I'm exact opposite of you. I will look at what there is on the first page or the other extreme would be I'm looking for something that's specific and then I'll do a keyword. Okay. Not anything in between. Okay. I'm gonna say, yeah? It's my 10th that's a Serendipity. So I open up a new browser window. I look through 100 or 150 tweets. I open up the ones that are good and other tabs. If they turn out to be good and interesting, I repost them, I rarely retweet. I use via a lot. Okay. Because I often don't. What is that? Just a curiosity. Do you retweet post versus retweet? Rather than just saying Donnie said this and just hitting the button to retweet it or retweet Donnie and do it, I will often put my own 100 character headline or something but I'll credit Donnie for putting it up there by saying a via. So it's my way of having an influence in the conversation rather than just. And usually I want to change the headline. Do people use hashtags? Yeah. Yeah. Yeah. I'm basically saying that the pattern is the same as Ethan's except I retweet Donnie's place here. Yeah. And I'm more even back on it. Interesting. Interesting. So it's weird like I look into Twitter far more than I do on Facebook just because they overhead just feels, it feels faster somehow. It seems like I get to the gist of it much quicker. But if you look at the number of Facebook status updates versus Twitter, it's orders of magnitude more. There's many more people using Facebook right now than Twitter. I don't know how that's changing. Donnie? To combine Facebook and the Twitter, for example, for me I have both of these two accounts and I can suggest to the users to log on at the same time with these Facebook accounts or Twitter accounts. And for the research perspective, you can compare what has happened and for the users, it can arrange their friends together. Is that possible? If people using the Facebook API, you can combine the two. You cannot get into the inbox, which we use some of the inbox parameters to build our model. But are you suggesting that we take information from Facebook and Twitter to make a more holistic model? Is that possible? That is possible, yes. Yeah. If people would have to put in both logins though, so there's issues of trust and other interface issues, but it is possible to build a model based on both behaviors, yes. And another maybe different from you, some of you is that because Twitter is blocked in China. And Facebook is also blocked in China. So for example, this kind of result for me, it's not this. Yeah. It's just that those frequent Twitter's appears in the inner circle. Yeah. But all the circle is just same to those not frequent, almost the same. So I'm wondering to that these two students, these two can be used as to compare the also the same topic to isolate it, different kind of web or internet. Yeah. Is that possible? To a degree. For example. To a degree. I set an application to some SNS websites in China, used almost the same kind of application and compare the same one. We can look at some kind of the users and compare the same two topics and compare what kind of thing. You could do something like that. I don't understand the specifics of what you're asking, but are you saying like look at your ties with people on two different sites? Yeah. Yeah, that you could do. That you could do. People may, when they want to cross the war, they actually have maintained two kind of society. This is fascinating to me because. Communities. Yeah. If we can do some research on these kind of aspects to look at those people who have to cross the war. Yeah, yeah. And what are their communities? Yeah. How can they, they can combine together all of the relationships. Yeah, I hadn't even thought of that. For me, the wall is an ocean. Yeah. And I just, it's like my Greek family is just not on the radar. So it's interesting to think of it in terms of having active conversations, but just in two distinct spaces. Yeah. Yeah, it's not only the censorship thing, but also perhaps the different countries or different. Yeah. So and you're, did you look at the communities that built for you? Did they make any sense? Until now, the workman battles, it makes sense, but others does not. No, okay, okay. Yes. And or try using your interface to look at email. Cause I was thinking that you would have the content of the email perhaps to you as well as, you know, the, how many emails we send between different people, et cetera, et cetera. And that, that might be your environment. Yeah. And then also I was curious if, if you took people who had been through a disaster or somehow tied it, like let's say Mumbai, where people Twitter, et cetera, and tried to use the interface to look at some of the, the connections of people and how the information got out. No, we have not done either of those things. Or let's say earthquakes and the same kind of thing. Microsoft and IBM had built different email clients to address email from a number of conversations perspective. We have not done that with this tool. As far as the disaster relief, I think it's a fascinating idea. There's a lot of really good work that's gone on with the Red River Floodings in Minnesota and South Dakota where they went in and looked at how people actually use Twitter for the response. They went in and coded, hand coded millions and millions of tweets. We, if we can get our hands on some data. So I thought that might be something that is definitely worth looking into, but we have not done that yet. No. Mumbai, my friend. I'm sorry? Mumbai? Mumbai, okay. They looked at certain organizations looked at it from a security perspective. Yeah, I think they did. We'll hook you immensely on that. Spushehidi's new project, Swift River. Swift River. Which is basically designed to figure out how do you deal with the flood of data that comes out in the wake of a natural disaster and sift through it. They're interested in essentially quality assurance. They're basically, they're a distributive reporting platform. So their question is how do you get through this and get to the reliable accounts? Because one of the things that we're all discovering is that disaster tweeting is 95 to 99% retweeting. It looks mostly like there's a little bit of information that comes out of them and everyone just sort of amplifies again and again and again. Al Jazeera just published this wonderful provocative analysis suggesting that there were actually 60 Twitterers in Iran during the Iranian Revolution and that everyone else was just sort of Americans patting themselves on the back and saying, you know, look, we're having a revolution out here. But they're doing massive data collection and analysis on this and it would be a fun test to know if you want to work specifically on that. They've done like an analysis and information flow to see who people choose to tweet versus. They're building infrastructure. So their theory is that the next disaster is already around the corner. People just grab their tools now in the disaster. So, you know, just as soon as they'd finished sort of focused on Haiti, someone grabbed it and built the platform for Chile. So they're building pipes. They're not social network analysts but they're really lovely open source people. And if you wanted to come in and do analysis on top of it, that's an easy introduction. That's a good idea. I was just intrigued when you had your great question mark slide up and you said that mapping out your relationships like that, you thought violated the Facebook Terms of Service? Yes. How so? They have, you know, you're not supposed to store data past an X number of hours from Facebook. And that has been stored in my computer for several months now, almost a year. So, Facebook has very specific terms of service for how you use the site, for how long you can store information afterwards. And that said, more than 70% of Facebook apps probably do violate Terms of Service. But that's an interesting discussion I wanted to bring up because it's something I would like to have with members of this community in terms of how to actually do research and analyze data in the most ethical fashion possible. Yes? With journalism coming at this from the position of a journalist. I was reminded, the NPR did a story and I recently talked to a radio journalist who was among five people who went into a farmhouse in France. I'm sure many of you have heard about this experiment and they cut themselves off from the world except for Twitter and Facebook. Okay. So that they could think through how they saw the world if the only way they saw it was through their Twitter feeds and through Facebook. And they've been writing and thinking and they're doing a lot more analysis and trying to think about how it relates to them what their job is. Yeah, yeah. And I think the word that you hear about journalism today is trust, either lack of trust or how do you regain trust? And I'm wondering if you thought at all about how thinking of journalism maybe is news and information. How, whether you can kind of track that within this in terms of how people gain trust to move from an outer circle to an inner circle from that perspective. That's an excellent question because that kept coming up in like the urban rural studies and Nancy Baim actually did some great work with one of her PhD students in interviewing people in rural areas trying to figure out what they wanted more to participate more with social media. So we got very lucky in that her work was going at the same time as ours and we could talk to them about it. The trust problem is really hard. There's different research going on with trust on video and audio. Trust with text, the literature that I've seen just hasn't been very conclusive about it. And the thing with tweets is that at 140 characters, you don't get a lot of the context that you normally get and the cues that you would use to actually establish, yeah. But what people do use there is they use reputation quite a bit. So in terms of reputation systems, that's probably where you're gonna get at some of the more trust issues. So of these people who put themselves into the farmhouse, did they know each other beforehand or are they strangers? They knew each other through the journalism community. They were from different countries. One was from Canada, one was from Belgium, I believe, one was from Switzerland. I mean, they were, you know, and they knew each other to plan this experiment and viewed it as an experiment. So now there's gonna be kind of thinking that develops out of this, that they try to put into their work once they've emerged from this. Yeah, so where are they coming out? This happened first week in February, so it's a fairly new. Okay, okay. Yeah, I wasn't aware of this. I'm looking forward to see what they have to say when they do. One of my colleagues actually did a study where he put in information foraging into clustered nuns to see what their experience is. But in terms of, the trust question is excellent. I wish I had a better answer for you. That's a cool, just a thing. Yeah. I think trust has essentially changed with the proliferation of weak ties in social networks. So, you know, the $100,000 is one example, but you know, I'd ask my father for $100,000 but for the best restaurants in lower Manhattan, I would be. But the reason it's also changing is historically in the sociological definition of tie strength, a strong tie implies trust. But one example of where, you know, trust, you can get trust in the weak ties. For example, let's say you walk into the street and there's a police officer. I mean, they're probably a weak tie, but you do trust them as well. You trust a fireman. You have a lot more of these, like different types of relationships on Twitter than you do in the face-to-face world. And it's easier to separate them face-to-face than it is on Twitter. So someone that went to the same school with you, do you trust them or do you not trust them? Someone that, you know, is in your church, do you trust them or do you not trust them? So it gets even murkier in Twitter than it is in face-to-face. What are those text-based signifiers you can add? Doesn't that, I think I'm about to make a very obvious point, doesn't that suggest that the weak, strong tie polarity just isn't all that useful in an online environment? Which also suggests that maybe it's tied to the graph pool. I think one thing it implies is that trust isn't mapped to strong ties, the way people thought. I mean, that I believe is true. I believe that there might be several types of strong and weak ties, maybe not just the one magical strong and weak tie that we keep talking about. Regardless, people will take weighted sums of all of them to come up with one universal number, but. Yeah, I also would like to support David's point because in my reading of Grand Ovidder's paper on strong and weak ties, what he was really interested in was heterogeneous versus homogeneous ties. So what he's interested in is are your ties all the same or are they different? It was a side effect of it being strong and weak. If you happen for some other reason to have really a bunch of heterogeneous strong ties, it'd be hard to have, that would be great. And you have lots of weak ties and they're all homogeneous, it doesn't do you any good at all. So the focus ended up being on something measurable, but it wasn't even the point of his paper. And so I think for a lot of things, it is really important to figure out if that's what, both that it's maybe too binary, but a lot of times it's not even the thing you wanna look at. And I think that was a very influential paper, but it should have been strength of heterogeneous ties. And it said that, well, one promise if you want a lot of heterogeneous ties, you probably are also end up with a lot of weak ties. So he did, I think a lot of people took that from that paper and that's why so much work after that was in homophily and in terms of similarity as opposed to differences. There's been a lot of work on homophily especially in the health environment, here at Harvard actually some of the best workers going on in that domain. I guess because so much of that work was going on, we weren't as interested in homophily as we were in, our ultimate goal is information flow. So seeing how that happens and seeing how these bonds, I'm personally interested in how these bonds evolve over time. I don't wanna, I love this work, I think it's great work. I'm curious that because this model worked in 2008, when new technologies come along, what we do is we appropriate our face-to-face behavior onto any new social media that comes along. But if you're interested, for instance, in homophily, one question is like, if you did something, if you look at, keep your, the wee metal interface and say, okay, here's a measure where I wanna look at how similar is what this person tweets to what I tweet. So, I mean, that might tell you other things. And so if you're looking for similarities, that might be the pieces to be measuring instead of trying to get at those issues through strength and weakness. So what do you think about this, to be within the application that you're creating? Take a look at some of what John Kelly's work is right now, clustering blogs around, not how they link to one another, but how they link to third parties. And so he's essentially taken large sets of blogs. His best-made work is on the Iranian blog is here. That's sort of more impressive because he doesn't speak Farsi. So grab a bunch of Iranian blogs, ignore the blog roles, because that's got all sorts of social artifact associated with it. Look to see what everyone links to. Don't really look at the links to blogs. Look at those third-party links. And he was able to tease out four major and sort of 10 minor clusters. It's work that he did here with Bruce Upline. And it's really gorgeous stuff. You've got all that data. It's a little tricky because everyone uses all these URL shorteners. But if you grab all those links that everyone's linking to, you could imagine another way of clustering people within this based on sort of what information people have or don't have coming out of them. That could be a fun experiment. Well, there's a lot of cool work going on about who's important in the network in terms of information disposal. Up until recently, people kept saying that you want to have high centrality in many, many connections. And it's been some nice cool work recently talking about how it's not just that, but you want to have a large mass of people that are easily influenced and will believe just about anything that they hear or see. And that will get stuff spread through the network much faster than one very central person. The homophily work is fascinating. I think one of the reasons that we sidestepped it a bit was because that we wanted to look at it from just a different perspective. There's so much work with homophily. We felt that the field was just kind of saturated a bit with it and went, just try something new. I mean, no matter what you do, it still comes down to that. And not just homophily, but what interests me more than homophily right now is sort of like complementariness. People always group people by similarity as opposed to someone that might actually provide something different. And that's somewhat what a strong tie brings to the table. So when they look at like, people have looked at strong ties and weak ties in terms of creativity of groups and a lot of that work's been done in business schools. And so looking at not just similarity, I think is just for me, interesting. Wendy, you had your hand up, I don't. I'm thinking about this in terms of in connection to Helen Nissenbaum's work on privacy and contextual integrity. And it struck me that your tool is providing contexts. And so it would be really interesting to get data about things like what people call their groups. Do you start seeing the same terms recurring? And if you add interface to allow people to move members around, do you see that there's a definite sense of where somebody belongs? And it's clear that person was just miscategorized like the X you referred to. But in terms of privacy, you could take it a step further. Like one thing we haven't encountered yet was what if people start publishing their inner circles and someone's like, why am I not in your inner circle? And they really want to be. Or they feel like, why is he or she in the inner circle? And so that's something we haven't encountered just yet. But from the privacy perspective, I would love to talk about this some more because as soon as you start, one of the dangers, like there's this beauty to tie strength when it was nice and ambiguous, as soon as you start putting a floating point number to it, things start to change. And it loses some of that beautiful ambiguity. But with interfaces, you can still keep some of it as long as you don't make everything just very explicit. So there's this, one of my biggest fears with this work is I don't want to destroy any relationships. And because at the end of the day, one of my first quotes was that what attracts people most is other people. And that's one of the beauties of Twitter. That's one of the beauties of IRC. And we don't wanna, we wanna make sure that, like I'm in the computer science department. We are builders. We like to understand the social forces and the theory behind our work. We also wanna see how we can actually encourage different styles of relationships and hopefully not destroy them. One of the things that's always happening in real life relationships is that they're coming from the sociology literature on social devices. It's always being negotiated. Status, your relationship, always under negotiation. So one of the problems with placing people into these sort of fixed circles is that you lose the power to continually negotiate the relationship. And if they stop being very helpful, you wanna be able to jettism them from your inner circle before they move away. For many reasons you have an argument and they do something you were not expecting. You need some of that. Yeah, so one thing that's interesting about some of the algorithms, the way we envision using them is that they do learn in real time. So based on interactions, so like, so if Ethan tweeted me, so we would get slightly stronger in that realm. If something were to happen and I were to remove him from my, that would be incorporated to the list. But it would be sort of, what's interesting there is the explicit cues versus the communication cues. Like you can bring in cues from personal experience and say, look, this is what I'm doing now versus what happens from the dynamics of the conversation. So looking at those as well. And again, those are what were interesting in the relationship scenarios where our algorithm broke. So I'm really intrigued in that right now and I'm also intrigued by the positive and the negative. So for example, in terms of political discourse, it turns out that if you really want someone to reply to you, being rude really helps in a political context. Whereas in a technical form, if you really want to get a reply, being polite helps. So this element of context is also key, which so far we've pitted as a black box. But in the political scenario, getting a reply back will probably give you a negative reply back. Most people don't reply, yes, I agree with you in those forums. They'll reply, no, but did you look at this or did you look at that? Blocks are different. Blocks tend to be more echo chambers in the political realm. But there's a lot of different features to look at, but I like the idea of looking at sort of, like external features that are not captured in Twitter. For example, most of our life does not happen in Twitter. It's good. Depends on who you are. True. We are. Oh.