>> Michael Gamon: Okay. It's my great pleasure today to introduce Justin Cranshaw from Carnegie Mellon. In the Mobile Commerce Lab. And he's done -- most people actually know about this already, but he's done some pretty amazing work in this livehoods project using social media to provide information about cities and how people behave in cities. And I know that ties in very nicely with some of the research that's going on here. So it's really nice to see Justin here. And we also, as we saw before, they have some opportunity to explore Seattle through the livehoods lens. Thanks. >> Justin Cranshaw: Thank you. Thank you. So thank you, Michael for the invitation to talk here. It's my pleasure to join you all. So, yeah, I'm going to talk about this project, the livehoods project. But kind of the -- the thrust of the talk is utilizing social media to understand the dynamics of the city. So this is work, as Michael mentioned, done in the Mobile Commerce Lab. And it's joint work with Raz Schwartz, Jason Hong, and my advisor Norman Sadeh. Now, I understand Jason Hong spoke here a few weeks ago. Was anyone in that talk? A couple. So I don't think there will be a huge overlap. But I think it should complement it nicely. So to give first -- to set some proper expectations I think this talk is going to have a rather large and broad scope, perhaps more broad than you're used to. In addition to covering our research on the livehoods project, since the general research area that we're working in is quite new, I'm going to be successful motivating the work with background from urban studies, including works from sociology, social psychology, and urban design. And the kinds of main contribution and the take away that I'd like you to get is that we're kind of introducing this new methodology for studying the dynamics and the structure of cities. And I'd like to -- for that to be the main take away is this methodology, this new approach. So that being said, this talk is about the convergence of three concepts, cities, social media, or social computing and ubiquitous computing. In particular, I'll be talking about how the latter two concepts, social media and ubiquitous computing are altering the dynamics of cities but also allowing us to peer into the inner workings of cities at a level of detail we haven't been able to do before. So I tend to think of the synthesis of these three concepts as falling under the umbrella of this new field called urban computing. This is a field that I guess like other new fields has lots of different names. Some people call it urban analytics. Others call it urban informatics. IBM loves the term smart cities for some reason. And architects tend to love sentient cities, which is a little too sci-fi for me. I don't know why a field with so few researchers needs so many names but such is life. And but broadly speaking this field seeks to utilizing computing technology and typically pervasive computing technology, to enhance the efficiency of cities and the general overall well-being of the people that live in them. So you might ask, okay, well, that sounds nice, but why cities, right? So smartphones and ubiquitous computing, they're used everywhere. They're used in villages, they're used in small towns, they're used in suburbs. And additionally social media is almost as widespread as the Internet itself. So what's with this focus on cities? So the question arises why this focus on urbanism. And to address that, I'm going to talk about why cities are different from suburbs, from towns, and from villages. So that's a really complex question. And I'd love to spend two hours talking to everyone about that. But I did promise to get to some computer science as well. So I'll talk about two reasons why cities are different. And the first is in fact density. So density is actually the most critical reason why cities are different. In particular, because density is sort of a defying property of cities. In fact, if you were to define what a city is, probably density would come into that definition. And so one of my -- my favorite quotes from Sherlock Holmes, homes says to Watson the city is full of those whimsical little incidents which happen when you have four million human beings jostling each other within the space of a few square miles. He says amid the action and reaction of so dense a swarm of humanity, every possible combination of events may be expected to take place. That's from the Blue Carbuncle if you're a Holmes fan. And I think one of the truly great promises of urban computing is in exposing, organizing and harnessing this dense swarm of humanity. And so I think that's pretty exciting. But this density, it creates a sort of nexus of resources, and of people and of ideas that result in a sort of cosmopolitanism that was described by Lewis Mumford as the think in essence multiplies man's power to think, to remember, to educate, and to communicate. So it's this sheer density of people and ideas and interactions of the city that presents an enormous array of opportunities for innovation through persuasive technologies. Now, I want to stress that I'm not implying that density is always a good thing for cities. In fact, it's not. So -- but such a kind of nuance discussion is beyond the scope of this talk. I'm only implying the rather self-evident and noncontroversial claim that density is a defining property of cities and it's why we're looking at cities through, in essence, ubiquitous computing. So a second reason why cities are different from suburbs, from towns, and from villages is that cities exhibit an immense amount of what I call sociocultural spatial variability. Now, another more common word for that is cities have neighborhoods. So I'm going to -- this is a concept I'm going to get back to later in the talk. But something that I just wanted to plant in your heads for a while to think about. But neighbors are quite important for cities. And actually it's kind of an emergent property because of the density and complexity of cities. So traveling small distance in cities is quite difficult. And so distance is quite important. And that's kind of why these -- this variability emerges. So okay. I told you one of the take aways I want you to get is that we're introducing a new methodology for studying cities. So if that's the case, it's important to understand how researchers and scholars have traditionally studied the urban landscape. And so I want to briefly overview, in essence some traditional methodologies for studying cities. So one approach, and I mentioned him earlier, is one that was taken by Lewis Mumford. So in order to better guide the potential forum of future cities, Mumford studied cities through the history of how their form evolved. So he wrote one of the most influential works of scholarship on this subject. This is a book entitled The City in History. And he really looks at, like, architectural evidence of how cities evolved and how the functions that they support evolve through time. And his motivation essentially is that by studying the history of how these forms evolve we can unlock greater potential for the future forms of cities. Another approach was taken by a guy named Stanley Milgram. You probably heard of this guy if you've taken like undergrad psychology course. But he's famous for things quite different than this work that he did here. So Milgram -- does anyone raise their hand? What's Milgram famous for? >>: [inaudible] experiments. >> Justin Cranshaw: Yeah. So the electric shock experiment is one. Another one is the 6 degrees. He did that study. So those are the two main experiments that he's famous for. But he also when he was at CUNY in New York, he embarked on an amazing program to study the psychology of what it's like to live in a city. And so he did these really fascinating studies of how the urban environment affects the individual psychology. So one example, a famous example, actually, he wanted to study how social norms affect the process of riding the subway. And so he conducted this experiment where he had able bodied graduate students enter the subway and ask people who were sitting to give up their seat with no reason given whatsoever. Right? So they just repeatedly had to ask these people, can I take your seat. And so if you've ever lived in New York or spent any time in New York, this is a pretty strong social norm. You just don't do that. And so the purpose of this experiment was to kind of record these people's reactions to the question. But, in fact, the interesting outcome turned out to be the effect on the experimenter of actually repeatedly having to violate this norm over and over again. And so they actually experienced some physical illness because of having to violate this norm. I mean, it's such a strong, powerful force that violating it was very difficult. So his work stands out as testament of actually the power of social norms in cities. Another example was done by a guy named William Whyte. He did some amazing work where he spent thousands of hours recording, taking pictures of, and observing people in New York just interacting in public spaces. Nothing more. So just trying to understand how the design of the environment and the various factors, for instance the -- what time of day it is, where the light is shining, how these affect how people utilize public spaces in the city. And there's some fascinating film clips associated with his work, the life of small -- the social life of small urban spaces. Whyte's work resulted in a better understanding of how people naturally utilize public space in the city as a function of things like urban design, light, and time of day. Another example is more of a sociological study. There's a sociologist from Chicago named Gerald Suttles who actually conducted some embedded observations. So he wanted to study the social order of life in Chicago slums. And so he spent three years living in a tenement in Chicago and actually just documenting what life is like there. It's a pretty amazing study to read actually. And finally, Jane Jacobs -- I don't know if anyone has heard of Jane Jacobs. But she wrote one of the most influential works in urban design, The Death and Life of Great American Cities. She was a writer and an urbanist. And her work essentially was a sharp critique of existing theories and planning. And she argued vehemently against rigid planning movement that is were popular at the time. So there was this great movement in the first half of the last century where urban designers thought that they could harness and control the chaos of the city. And so you had these really like spatially geometric plans that looked great on paper but in implementation they hardly ever worked out right. And Jane Jacobs was a strong opponent of such trends. And her -- her -- she preferred a more market driven approach to urban design where the cities would grow organically based on the demands of the people. Essentially her view was that the people can determine what the best urban design is for them. And why am I talking about all this? So why should you care? What I want to emphasize is that studying cities is complex. All of these studies that I mentioned are really intricate, long-term ambitious studies. And it's actually very difficult to actually answer any questions which any rigor about a city. So it requires historically immense hours of field work, interviews, surveys, observations. And even though the works that I listed just now have been hugely influential, such methodologies can only ever offer a small-scale view of the city. And they're inherently partial, and they're biased, right? I mean, these are essentially one individual studying one particular aspect of the city. And by virtue of that alone, these are biased works. You know, they're powerful works, they're good works, but they are biased. And we should recognize that. So can we do better is the question. Can we come up with a large scale away of studying the city? Can we come up with a less biased way of studying the city? And can we come up with near realtime methodologies for studying the city? So that's kind of the main goal for the rest of the talk is developing that. So that leads into this next topic of computing in the city. Now, it should not be a surprise that technological revolutions have profound effects on the urban form. To take this point to a logical extreme, for example the technological development of farming an animal husbandry led to the change of urban forum, if you want to call it that, from paleolithic cave life to neolithic villages. And then to take the next step, the development of masonry and bronze weaponry led from us these neolithic villages to actually neolithic walled cities. So, yeah, I mean, that's -- that's kind of a distant example. How can we actually look at how computing currently is changing is urban forum? Now, I -- I say computing, but I actually specifically mean ubiquitous computing because I think that's having the greatest effect on the city and the urban forum. So what is UbiComp? Ubiquitous computing aims to move computers off of desktops and essentially into the environment to a point where they're actually invisible, they're so ubiquitous, they're so blended into the environment that computing actually disappears. So smartphones actually represent the first kind of wave in realization of this vision. I wouldn't say that smartphones are ubiquitous computing, but it's kind of the first step to getting there. And it has been -- they have been obviously amazingly successful. I don't need to tell any of you that. But you have these devices that are permeating our daily lives and permeating -- and changing how we interact with each other and with the environment around us. And so you also have these kind of new and rich sources of social data that are expanding our understanding of the city and creating new possibilities for interactions. So I have this diagram where you have these two directions, right? So you have on the right hand these devices are changing how we interact with each other and the environment, the city. And on the other hand, using these devices and using the applications that are embedded on them it's actually creating data that helps us better understand the city. It helps us better understand the environment. And so you have these two forces that are kinds of converging. Yeah? >>: So it's [inaudible] sometimes about ubiquitous social connection is that it's -- what's the word? It's sort of metaspatial or something like that. Like we can connect with people regardless of geography. >> Justin Cranshaw: Yeah. >>: Right? So I don't know. I mean, what are your thoughts? >> Justin Cranshaw: I think that's definitely true. But I would argue that it's not -- you can connect with people regardless of geography. But distance will always matter. Right? And because distance will always matter, what's happening is you have a change of what distance means. You're almost changing the topology of the city, right? So things are still spatial. But spatial means something slightly different now. So if like, if for instance I'm using foursquare and I'm watching my friends check into different places in the city, that provides this kind of ambient update about places that I might want to go to. Right? So it's changing my understanding of a spatial region. But it's still spatial. All right? So distance still matters. I think it's a function of the fact that -- I mean, let's foursquare for example, right? You don't get updates from all of your friends, right? Because that's too noisy. You don't care about all of that. You only get updates from people that are nearby you. >>: [inaudible]. >> Justin Cranshaw: It's not entirely true, but it is by default true, right. So you get these push updates, right, when people check in. >>: Yeah. But it is sent to like who you define -- I mean, there's another type of distance. But like the way foursquare updates [inaudible] on my phone are who are closest to me in person space, not in physical space. So almost all the updates I get while I'm here are from people home in New York. >> Justin Cranshaw: Okay. So mine is set to physical space. Maybe it's a default or something. But in any case, I think my point is still valid that distance does still matter. And it matters because we're still ultimately interacting with physical things, right? >>: [inaudible]. >>: Yeah. That may be. But it may be ->> Justin Cranshaw: So what if -[brief talking over]. >>: Physical space relative to your hometown. >>: Right. >> Justin Cranshaw: So what it actually does for me is the last place that I checked in, I will get updates from people nearby. So I still get updates from people in Pittsburgh because last place I checked in is Pittsburgh. >>: But you could also turn on updates from anyone. >> Justin Cranshaw: Yeah. >>: So like I get updates from all my friends. >> Justin Cranshaw: But in any case, they have some default that takes distance into account. And I think that's -- that's a critical realization that distance does matter and it will always matter: And going back to what I said before, because of the complexity of the city, going small distance in Seattle and any big city is difficult, right? And that will never change. You know, we can get new methods of transportation, et cetera. But that will never change. >>: And I think changing the nature of what script geography means to people [inaudible] is cool. I mean, it's interesting. >> Justin Cranshaw: So, I mean, that leads us very nicely to these locations based social applications. So thank you for that question. So one example is foursquare. I don't really need to go into what foursquare is but essentially you check into a venue and that venue is shared with some subset of your friends or your social connections on foursquare. And a check-in is typically given -- or it typically includes a wealth of information about that case. So you'll get the category of the place. You'll get various tips about the particular venue. And you'll get things like interesting things that you can do there. Another example is a status update like a tweet. Tweets can contain mentions of various place entities. So in this tweet I'm mentioning Belltown, I'm mentioning Seattle, and I'm mentioning a winery. These are entities. These are entities that are associated with some location. But in addition to that the tweet itself can be of course geotagged. And so I could have the location that I am currently at. Another example of such media is a photograph, like Instagram. Most Instagram photos are geotagged. I think you can turn this off. But I think by default they're geotagged. And the user can manually elect to tag that photo with an actual named venue from the foursquare database. So you have both the latitude and longitude, but also the venue of the place. This is a new -- one that was new to me. Has anyone seen this before? Forkly? This is great because it's an idea that I -- I wanted -- I thought would be a great idea. And it turns out someone is doing it. But what it is, it's essentially foursquare or Instagram for meals. So you -- if you're out at a restaurant and you're enjoying a greatly meal you take a picture of it, you describe it, and you link it to the venue you're currently at. But in essence, I mean, this is all -- some people like to use that term Internet of things. I don't know if I like that term. But it's essentially putting a virtual entity about some physical object, in some case a meal, in some database that's online and available for other people to access. So again I want to go back to these two points. I think these technologies are changing how we interact with each other and the world around us. But they're also generating massive amounts of data about ourselves, as well as how we interact with each other and the world around us. So that leads me to this idea of studying the city with smartphones and with sensors. So studying the city through direct observation, as I mentioned, historically required an immense amount of work, field work, interviews. And these are kind of small-scale partial studies. You get really great insight from these. I don't want to discount these methodologies. I think you get really amazing insights from them. But again it's a partial view. And so what we seek is computational ways of uncovering things like local cultural knowledge by leveraging these rich new sources of social media. That takes us to our work on the life -- yeah? >>: [inaudible] you do recognize that you've been looking at social data [inaudible]. >> Justin Cranshaw: Yeah, I mean, there's so many biases in social data. >>: I just -- I wasn't ->> Justin Cranshaw: And I'll get to that at the end. But it's an excellent point. >>: So when you talk a lot about the cultural data, especially like in Chicago, there's rich cultural neighborhood of inaccess to the Internet. >> Justin Cranshaw: Absolutely, yes. >>: You do lose that snapshot. >> Justin Cranshaw: And that's a really big problem. And, I mean, I can talk a little bit about it now because you brought it up, but, I mean, I think this type of analysis is thinking forward, right? So, I mean, if you think of for example Facebook, right, Facebook Five Years ago had really -- even then had pretty strong demographic biases, whereas now these demographic biases are much less. Still you get people using the service differently than others in different cultures. So you'll always have the kind of variability but you lose some things like sample bias. I think that gets less and less of a problem. Currently sample bias is a big problem in these kind of mobile social data. You know, it's typically mostly wealthy people. They're biases along almost any demographic you could mention. Yeah. Thanks. So I want to return to this idea of neighborhoods. So neighborhoods play a number of essential roles in the daily lives of city dwellers. So by virtue of our kind of everyone day experience with neighborhoods, I think most people have an intuitive grasp on what they are. But so you understand what the idea of a neighborhood is. But it can actually be pretty difficult to describe concretely what a neighborhood is. And let alone it can be -- if you wanted to actually with precision kind of draw the boundaries of what a neighborhood is, that's actually difficult thing to do. So, yeah, we can talk about a little bit about what roles neighborhoods play in a city. Essentially I talked about cities being complex and dense. Neighborhoods provide a sort of order to the chaos of the city. They help us determine things like where we should live, where we should work, where we should play. They provide things like a safe haven. Safety. They provide a sense of territory. They also serve critical roles in many municipal governments. They're not necessarily a political unit. In fact, they're rarely a political unit. But they do serve a sort of organizational purpose in terms of resource allocation. And that's typically because municipal governments interact with neighborhood groups and so the neighborhood groups have to communicate back and forth with municipal governments. They're also centers of commerce and centers of economic development. And interestingly enough, they're also brands, right? I mean, company -- businesses that are located in a particular neighborhood, they like to advertise that they're in this particular neighborhood, if it's a trendy place. And that's actually pretty important. And -- yeah? >>: [inaudible] neighborhoods, are they typically -- is there a pattern to how they develop like is there a certain percentage of a city's population or sort of -- you know what I mean? If you like city to city, would you see some sort of constant ->> Justin Cranshaw: I think you would see a pattern. But I think the pattern is -- it's a cultural thing, right? I mean, the development of a neighborhood is a socially constructed thing. It takes into account the history. It takes into account, you know, things like property values, cultural. And so I -- I think that is the pattern, that they're socially constructed. But can you specify that process? I don't know. That's hard. So the final point is that neighborhoods actually because of this fact that they are socially constructed, they provide a sense of cultural identity to the people that live there. You like to define yourself by the neighborhood that you live in. So I -- yeah, as I mentioned, neighborhoods are cultural entities. And because of that, we actually each carry around our own biases of how we view the city and its neighborhoods. So neighborhoods actually often evolve quite fast. And this can typically be much faster than the actual physical boundaries that delineate them. So one example in Seattle is South Lake Union, right? If you probably looked at a city map, A, that neighborhood might not exist. I'm not sure like what -- what the boundaries were 30, 40 years ago. But, these city maps, they tend to be drawn quiet irregularly. They're based off things like census tracks, and they kind of aggregate census tracks in a way that makes sense to whatever city planner is drawing the map. And these are done like every 30, 40 years. They're not done at the speed in which a neighborhood can actually come into existence. Even like thinking 30 years ago in Seattle I'm not sure that the landscape would look anything like it does now. And so these city boundaries, they can be misaligned with the cultural realities of the neighborhood. And that's where we're going to be getting into in this work. But to motivate the rest of the talk, what comes to mind when you picture your neighborhood? Just take like 10 seconds. So I don't know what it is, but I know the chances are you're probably not imagining this, which is a fixed set of boundaries on a map. Right? Chances are when you're picturing your neighborhood you're going to imagine something like this, right? What you think of is you think of the shops and you think of the sights and you think of the attractions and you think of the public art, and you think of the stores. Am I right or am I wrong? Yeah. How many people actually thought of the map? >>: Vaguely for a minute. >> Justin Cranshaw: Vaguely. [laughter] It's always like one or two. I gave this talk to a -- or a talk like this to a city -- a group of city planners and like half the room raised -yeah. But, yeah, so you think of these kind of cultural entities that define the neighborhood. And so one great quote that I think summarize this point is by Kevin Lynch, who is an urban designer from MIT. Kevin Lynch said: Every citizen has had long association was some part of his city, and his image is soaked in in memories and meaning. I think that's quite apt. So I mentioned Kevin Lynch because he did actually some amazing work to study this effect. What he did is he wanted to study people's mental maps of the city. And his kind of purpose was to try to understand the kind of legibility in the city as you said. He wanted to understand how the built environment affects people's perception of the city. So he wanted to understand things like basically knowing that each of us carries around this biassed views, he wanted to understand what modifies, what affects this bias. So what he did is he had people come in and draw their own mental map of the city. In his case, he did Boston. I think he did LA. And -- and Jersey City. I don't know why Jersey City, but he had people draw their mental maps of their city. And then what he did was aggregate these mental maps into one composite map of the city that showed basically the various kind of biases that people had, to have things like cultural attractions would be more prominent than kind of neighborhoods that people don't typically go to. This approach and this methodology was, again, repeated by Stanley Milgram, who I mentioned earlier. And Milgram wasn't necessarily interested in kind of this from an urban design perspective. He was interested in the psychology. So what is the underlying psychology that affects how people kind of navigate through an urban environment. But he did very much the same thing. He had people come in, draw maps of New York and Paris. And he'd give very little kind of rules as to what these maps should look like. And what they -- what these maps are, these cognitive maps, they provide a kind of deep sense of the local knowledge of the city, granted, biased by these individuals that are drawing them. But it provides a way of kind extracting what these biases are. And so I think it's a pretty powerful technique. And so that kind of highlight these two perspectives on the neighborhoods of the city. On the one hand you have what are politically constructed neighborhoods. As I mentioned, there's this process that comes in like 40-year cycles that defines neighborhoods. But on the other hand, you have these socially constructed ideas of what a neighborhood is, right? It's these cultural entities that define the neighborhood. And so I think each is equally important. So the politically constructed view is important from a municipal government to having something that's kind of static that they can kind of from year to year to year go back to. But the kind of socially constructed view is essentially how we every day interact with the city. Right? And so I -- what's nice is that our mental maps can shift as the city shifts. So as there's, you know, construction in South Lake Union I can kind of reconfigure my mental map to understand where I should go and navigate to get to places I need to go. So these cognitive maps, they can help shed light on this misalignment. And they can help identify he's distinctly characterized areas of the city. However, this technique of having people come in and draw their maps, again, it's one that does not necessarily scale. This is one you can maybe have a couple dozen people come in and draw a map for you, but then what do you do to it? Or do with it? So our ultimate goal in this project in the livehoods project is to find automated ways of discovering this local cultural knowledge and to actually be able to build these cognitive maps from data. And so can we discover these sort of organic views of a neighborhood? Can we extract local cultural knowledge from social media data? And can we by aggregating the activities of thousands of people discover a sort of collective cognitive map of the city? And so these are our research questions in this project. So the hypothesis here is that if we're talking about the character of a neighborhood or of an urban area, this character is defined not just by the types of places that are found there, but also by the people that make that area part of their daily routine. So if we can -- we can essentially characterize a place, namely by observing the people that are visiting it. And this should make sense, right? I mean, as you walk from one neighborhood to the next, you'll notice that the buildings change. You'll notice that the architecture change. You'll know that like the businesses change. But also the people change, right? I mean, these are cultural entities, right? I mentioned that people are defined -- their identity is linked in some way to a neighborhood that they live in. So we should be able to look at actually where people are going to kind of define these areas. And, again, there's another great Kevin Lynch quote: The moving elements of a city, and in particular the people and their activities, are as important as the stationary parts. I think that's great. So to discover these areas of unique character, we should look for clusters of nearby venues that are visited in essence by the same people. So that brings us to this idea of trying to cluster the city. So if neighborhoods are a manifestations of a sort of sociocultural spatial variability then we should be able to detect this variability through clustering. So what properties of these clusters should we be looking for? Well, first they should be geographically contiguous, right? I mention -- or maybe I didn't, but the data that we're working with is going to be foursquare data. So we'd like to find clusters of geographically contiguous foursquare data. Clusters each should have a distinction character from one another. And this should be perceivable by city residents. The clusters should be such that venues within a cluster are more likely to be visited by the same users than venues in different clusters. So what's the intuition here? So if you look at foursquare venues, these are the dots. And you look at people checking into them over time, you'll notice some patterns, right? You'll notice that like the green and the orange guy tend to stay on the left-hand side of the map and the purple and the pink guy tend to stay on the right-hand side of the map. This is what I was describing before, is that people tend to stay in their neighborhood, in their kind of cognitive map of what their neighborhood is. And so we'd like to kind of look at these patterns and extract what these organic neighborhoods are. And so if you watch these check-ins over time, you'll notice that groups of like-minded people, they tend to stay in the same areas. So we go aggregate these behaviors, these check-in activities, and determine relationships between venues. And then we can use these relationships between venues and apply some standard graph clustering techniques to discover these actual -these unique -- these uniquely characterized urban areas. So that's -- that's the intuition of what we're doing. So we build some similarity and matrix between venues and then we look for clusters. The clusters we're calling livehoods. This is meant to kind of reflect the kind of dynamic nature of the way they're constructed. So what's the actual methodology that we're using? First what we need is a kind of notion of venue-to-venue similarity. So we can get this actually by looking at the check-ins to each venue. So in essence what we're doing is we're treating a venue -we're modeling it as sort of a bag of the people that have checked into it. So rather than like a document as a bag-of-words, we think of a venue as a bag of check-ins. And so given this kind of vector representation of a venue we can look at things like any sort of vector similarity between venues. In this case, we're using cosine similarity. And that's, in essence, how we measure the similarity between venues. But there are some problems with this approach. First when you have check-ins and venues it turns out that the resulting similarity graph is quite sparse. All right? So if you have something like a hundred thousand venues in the city, there's like, you know, a billion possible edges and you only have something like a million check-ins for this city, right? So you're not going to get a dense observation of all kind of pair-wise relationships. That's kind of natural. The other problem with this approach is that in a city similarity tends to be dominated by what I call hub venues. So like in New York, for example, Grand Central Terminal is a hub venue because everybody's been there, as is kind of the airport and various kind of venues that are kind of dominated by everyone in the city. And so that really kind of skews the topology of the graph. Because these -- these pair-wise venues will have high similarity. And they could be pretty far apart. And so that's going to kind of collapse the graph towards these two venues. So we need to kind of overcome these two problems. So in addition to capturing the social similarity between places, we also want some notion of geographic contiguity. And to do that, what we do is simply add a constraint to the graph. And that has the effect of both adding this geographic contiguity, but it also helps us with this kind of hub problem. And so what we do is we connect each venue to its, say, M nearest neighbors and we only look at similarities along these connections. So if it's not one of its M nearest neighbors by geographic distance we kind of ignore any kind of social similarity that could exist. So that solves this kind of hub problem because you're -- you're not going to measure the similarity between Grand Central Terminal and JFK. The other thing that we do is that we add some small constant among these nearest neighbors. So we'll actually -- the way then the similarity graph will be the social similarity that we measure between venues plus some small constant. And basically what this does is it -- if the similarity is zero, meaning you've never observed any sort of co-occurrence of users at these two places, then you're still going to take into account some of the geographic similarity between them. And so what you have is sort of you're linking nearby places but then you're kind of pulling them together based on the data that you do have. So that's our basic approach. And then we kind of -- you can think of this all as a big graph, we're connecting each venue to its M nearest neighbors and weighing it by some social similarity plus a small constant. Once we have this built, we conduct some very standard spectral clustering. We use the version of spectral clustering by Ng, Jordon, and Weiss. This works pretty well for our purposes, and it's nice in that it's well studied and easy to implement. We -- so, again, this is a parametric model. We kind of use the popular heuristics that kind of everyone else uses to pick the number of clusters. So we'll have some maximum allowable number of clusters and some minimum allowable number of clusters. And we'll select the K based on the largest kind of eigengap in that range. So I'm not going to go into much more detail around the methodology because we're using fairly standard approaches. Instead I want to talk about once we have these clusters from the spectral algorithms, we do some post processing to clean up the resulting clusters. So one thing if you've ever -- if you've done these kind of image segmentation before or other -- if you've used spectral clustering before, you'll notice sometimes you get kind of background clusters. So in image segmentation you'll get like a background noise cluster. And that's -- we get that as well. And so what we do to overcome that, is we -- if any cluster spends too large of a geographic distance, we'll actually split it up and we'll divide the venues to their nearest geographic distance. And then we'll also do things like split clusters into connected components. These are standard techniques as well. Finally once we have the clusters we do kind of a neat thing to measure the relationships between each cluster. And so what we do is we begin apply a sort of bag of check-ins similarity approach. But now rather than looking at the check-ins between any two venues, we look at all of the check-ins to each cluster, to each livehood. And so we'll compute a pair-wise similarity between all of the clusters that we've discovered. And the reason for doing this is that you can kind of review this related livehood of a way to sort of hierarchically group clusters. So once we divided and segmented the city, we can then see for a given livehood what are the other neighborhoods that people go to. I think this is actually one of the coolest things in the website which I will demo to you. Let me first talk about our data. So, again, we're using foursquare. But foursquare check-ins are by default private. But what we can do is gather check-ins that have been shared by people on Twitter publicly. And so we've done this. We're actually combining a dataset from other researchers of 11 million foursquare check-ins with our own dataset of approximately 7 million check-ins. And, yeah, so once we have these kind of Tweets of the check-in we then align the data with the underlying foursquare venue. Let me talk about the website a little. You guys all saw this when I came in, but essentially these dots are venues. And nearby dots of the same color form a livehood. So, again, the colors don't have any meaning. So if you have two clusters that are kind of far apart, they're not related in any way, think of it as more of a map coloring problem. So here is the livehood SoHo in New York. We also add things like various descriptives, statistics about the check-in of the venue. Here's an illustration of this related livehoods feature. So for here I'm settling SoHo, and I'm showing what livehoods are related to it. And I think this is kind of neat because it highlight various transportation corridors in the city and various patterns of behavior. So one cool example in Seattle is that if you click on Bellevue -- or Belltown, Belltown is related to a few nearby neighbors downtown, et cetera. But it's also related to these kind of wineries that are kind of near Redmond, which I find very unusual, but it also makes sense to my model of the type of person that lives in Belltown. So the question is we have these clusters, right? How are we actually going to evaluate them? So we started by clustering various cities. Since we lived in Pittsburgh, we naturally clustered Pittsburgh. So here's an example of the Pittsburgh clusters downtown. But once we have them, what do we do with them? I mean, you can try to come up with some sort of quantitative evaluation method but really what we're after here is we know that livehoods are different from neighborhoods, but we want to understand how they're different. In our evaluation we want to be able to characterize what livehoods are. For instance, do residence derive social meaning from the livehood mappings? Can livehoods help elucidate the various forces that shape and define the city? And so if this is our objective, any sort of quantitative measure that we can come up with to evaluate the quality of our clusters does not really capture that objective. At least none that I could come up with. I'd be open to any discussion about ideas that people had. But so in this case, quantitative methods are going to fall short. And so we default to a qualitative approach of evaluating this quantitative model. So that required us doing some field work. So to uncover what insights our model tells us about the social section texture of city life, the only way to effectively explore the relationships between our model and the social realities of citizens is by going out and talking to people, by learning about their lives and the way they perceive the urban landscape. And so our goal was to see how well our clustering actually maps on to their cognitive maps of the city. And so we conducted 27 interviews. Each interview was about an hour long of residents of the city of Pittsburgh. These residents were recruited through various social media campaigns. We kind of had neighborhood groups post our flier on their Facebook page, things like that. And we conducted a semi-structured interview protocol that explored the relationship between livehoods, municipal borders, and the participants' own perceptions of the city. The one requirement we had is the participants must have lived in the neighborhood for at least one year. The interview protocol as I mentioned, it was sort of semi-structured. But in essence we began with a discussion of their backgrounds and their relationship with their neighborhood. We asked them to draw their own sort of cognitive map of their neighborhood. And given this map, we asked them is there any shift in feel in the neighborhood? Are there any places where the character changes or there's a shift? Are there any places in the city where they can think of where the municipal borders are changing? And finally, after all of this, after we gathered all this data, we showed them the livehood clusters and asked for their feedback. And that's an portion thing in mentioning. So we're not top loading the interview in showing them the maps, we're gathering data first, then showing them the maps. And then, finally, we showed them the related livehoods feature and asked for feedback as well. So I'm going to skip a few slides and go to actually some interview results. The slides I skipped basically gives some intuition of how we structured the interview. But I don't want to bore you with that. So the interview results are actually kind of fascinating. So they paint a picture of how people view these different neighborhoods in the city. Now, first I want to ask -- I know a couple of you are from Pittsburgh, but is anyone like familiar with Pittsburgh? Yeah, just the few that I know. So that's okay. I mean, I think that you'll be able to recognize the patterns that are going on here in your own cities. So the first example I want to describe is of Shadyside and East Liberty. Now, Shadyside and East Liberty are actually two contiguous neighborhoods in Pittsburgh. Both of them are kind of east of downtown. But despite being contiguous, they're quite socioeconomically different. They're separated, for example, by train tracks. And the tale of these two neighborhoods is that Shadyside is thought of sort of the rich neighborhood in the city, with some upscale shopping. East Liberty is often characterized as sort of a poor part of town. It once -- used to be a major commercial area of the state, but it was sort of destroyed by de-urbanization and failed renewal projects. There's these train tracks that separate them. There's a new Whole Foods that was built about 10 years ago on East Liberty right on the border of the two and some recent development that includes the building of a pedestrian bridge that actually spans the train tracks. So let me describe this in detail. So Shadyside, again, it's described as upscale. It's -again, it's not a sort of homogenous neighborhood. On its western end, there are some large houses with high property values. And on the right, on the western -- on the eastern end, they're kind of more apartments for students and young professionals. There are within this neighborhood three commercial districts. There's this area on the western owned called Walnut Street which has lots of national brands, upscale shopping, upscale boutiques, some restaurants. It's thought of generally as conservative. Not politically, but the restaurants there are not necessarily experimental restaurants. It's very traditional. Ellsworth and Highland Ave are the other two commercial districts in the neighborhood. And they have very few, if any, nationally branded stores. They're much more independent, locally owned boutiques. Local eateries. And they generally attract a much younger crowd. Again, I mentioned the train tracks between them. There's also a kind of about us way that goes along the train tracks. Now, across these train tracks is East Liberty. As I mentioned, East Liberty was once a vibrant shopping area. This vibrancy was destroyed by some pretty bad planning decisions as well as some trends of de-urbanization in the '60s and '70s. Currently it's typically characterized as poor. There's a lot of vacant store fronts. The store fronts that the -- the nonvacant store fronts are the ones that typically cater to lower income neighborhoods -- or lower income clientele. But recently, and following the construction of this Whole Foods that I mentioned, there has been efforts to revitalize the neighborhood. Now, I say revitalize rather than gentrify because the forces that planned Pittsburgh are kind of different. Gentrification forces aren't necessarily -- they don't necessarily push people out of the city. And that's a function of the low property values. So the historic shopping areas are along Penn Ave, which is this kind of central one, and on Highland Ave and on Centre Ave. And again, most of the recent development activity is long Centre and typically the intersection of Highland and Centre. So I -- again, the development has followed the construction of this Whole Foods. So the Whole Foods was part of this large vision that was started by this company called Mosites Company in Pittsburgh. And they had a whole regional vision for this neighborhood. I mean, basically it was their plan to revitalize it. So in the words of Steve Mosites talking about -- oh, I'm sorry. This green part here is actually the pedestrian bridge. So the pedestrian bridge is connecting the Whole Foods to this kind of independent shopping area on Ellsworth and Highland. And so in the words of the planner, Steve Mosites, he said the purpose was to connect the three shopping districts. We wanted a pedestrian bridge because we wanted as many connections to East Side itself. So East Side itself is a shopping district. South Highland 10 years ago was just starting to come around, and now it has all these great things. And also you have Ellsworth that's right there. And so the vision for this region was to connect, to bridge these two shopping districts. So there is kind of my mental map of Shadyside. Now, here's actually what our livehoods clusters did to that region. So you see, if you remember the kind of borders that we drew, it's actually splitting Shadyside into two. And it's combining the eastern end of Shadyside with this area of East Liberty. So the clustering, again, it splits Shadyside into two. On the western end it combines the high end national brand shopping area with the kind of high end property valued housing. And then it combines the kind of student apartment area on the eastern end of Shadyside with this new revitalization area of East Liberty. >>: So [inaudible]. >> Justin Cranshaw: Yeah. That's exactly where the Whole Foods is. It's below the V in revitalization. Area of East Liberty. >>: So the Whole Food is like [inaudible]. >> Justin Cranshaw: Yeah. That's exactly where the Whole Foods is. It's below the V in revitalization. >>: [inaudible]. >> Justin Cranshaw: Google? >>: Yeah. >> Justin Cranshaw: No, Google is ->>: [inaudible]. >> Justin Cranshaw: So there's a Google Pittsburgh. That's kind of off the map to the east. That's a different neighborhood. So here's some labels to help you. But so you have the independent shops on the east, the kind of upscale boutiques, kind of stead shopping on the west and then this revitalization emanating from the big orange circle. So let me talk about our interview results. So here's Kelly, who is a 29 year old resident of Shadyside again describing this neighborhood. She says when you go to Walnut Street, that's where I often see an older demographic. You'll see women and men above the age of 50 walking around with shopping bags. I don't see that demographic on Ellsworth ever shopping around. So I would say that's a big difference. You are going to see an older, straight, richer people on Walnut, and you're going to see much more younger, indie looking people on Ellsworth. So that's -we saw that this is one example. We saw this over and over again in our interviews. And that's kind of validating this split between the western end of Shadyside and the eastern end of Shadyside. And it's essentially validating that there's different demographics that operate in these neighborhoods. So validation of the kind of combination of the eastern end of Shadyside with East Liberty, this is Allison, a 26 year old resident of Lawrenceville. She said: If I didn't know anything about East Liberty and Shadyside, I would say this area here is the same neighborhood, centered around Highland. So that's centered around in essence where that orange dot is. They just seem to be blending the two neighborhoods. They seem to be blending. And in just as far as the businesses that are there go. And another resident, Erin, said -- and this was after she had seen the map. So after she had seen our clustering. She said: That makes sense to me. I think at one point it was more walled off, talking about East Liberty. East Liberty was thought of -Shadyside was wealthy, East Liberty was poor. But now there's all these nice places in East Liberty, and there's some more diversity in this area, so it's becoming more of the same. And I think Shadyside is almost shrinking and you only have a few streets that are really only wealthy. And so all of this highlight that our clusters are good at identifying these kind of areas that are in flux. So we actually asked all of our participants what areas in the city are kind of in flux or blurry to you. And I think 85 percent of them said that this area was the one that was in flux. And this is one of the few areas where you can see actually the livehood spilling across the neighborhood border. So it's pretty interesting to see that. The take aways here are that you can actually visually observe the concerted effort of the Mosites companies and others to blend East Liberty with Shadyside in the creation of this new neighborhood that they call East Side, which is actually a blend of the two words. And you can actually observe that through the activity patterns of residents. And so the new emerging developments in East Liberty are developing a character that is more similar to the younger, more independent, eastern end of Shadyside than it is to the more older, wealthier, staid western end of Shadyside. And so livehoods can help identify these emerging trends. So I'm actually -- I have another case study that I'm going to kind of skip because my talk's running over, and I would like to kind of play with the maps and take questions and stuff. But I will hop over to some conclusions. So, I mean, we found these -- so I give you one example. We found these over and over again, these confirmations of these patterns, making sense to people, in our interviews. So there's strong support for the clusterings. So the interviews showed that residence found strong social meaning behind the livehoods. And we found that livehoods can help shed light on the various forces that actually shape people's behavior in the city. And these include things like demographics, economic factors and cultural perceptions. There are a bunch of limitations that I want to talk about. First, although most neighborhoods had real social meaning to the participants, no algorithm is perfect. There are certainly livehoods that didn't make sense. And so we document these in our paper. I didn't talk about them in the talk here. There are obvious biases using this kind of data, foursquare data. And we view this as a kind of limitation of the data, not necessarily our methodology. So we feel that as you get better and better and more representative data, you'll be able to produce more accurate images of the city. There is definitely experimenter bias associated with tuning the clustering. So by and large the clustering tended to be pretty stable to parameter selection. Generally the location of the clusters was always the same. But you'd have -- particularly depending on the number of clusters that you'd select, you'd have various clusters combining or splitting. So the sizes of them changed. But the low indications were pretty consistent. But this is an anecdotal evidence. I don't have a rigorous exploration of the parameter sensitivity. Again, because of the biases in the data you have some populations that are entirely left out. The ones that don't use foursquare or ones that don't have smartphones. These tend to be kind of older and much more wealthy people that are too busy for these kind of technologies and poor people that can't afford these technologies. And so you have these two end of the spectrum that you're not capturing. I also don't want to overemphasize the kind of sharp division that is are being drawn on the map. I think neighborhoods tend to blend into one another. And, in fact, you could come up with model that is are kind of more soft in their clustering. But unfortunately the visualization needs to be kind of hard. It's hard to come up with a soft transition without doing like gradients or something like that. And, again, this is not a comparative work. This is a qualitative work. So we're not claiming that our model is the best, just that it does a good job, and it produces things that are meaningful for people. And finally, thanks. I'll take as many questions as you have. And I'm happy to also play with the maps, if you want to look at them. [applause]. >> Justin Cranshaw: Yes? >>: So I actually have had a lot of fun drawing the mental models with people when I was a Peace Corps volunteer in Africa. And I only worked with women. But I'm absolutely convinced that the maps they drew were completely different than what would happen if we even included men in those groups. So we would see like where you would go to get firewood or where the water was, or where they would even go to like sell or buy food, which I don't think would happen. >> Justin Cranshaw: Yeah. >>: As much if men were in the group. >> Justin Cranshaw: Yeah. >>: So, I mean, this is a bias, but I think it's a pretty strong one. Do you have any sense of -- are there any groups that if you split them with this data you would get totally different maps? >> Justin Cranshaw: I think that's -- that's absolutely true. And it's not something I had thought about too much before. But in particular you could imagine with regard to time, right? If we had time elements to our clusters you would definitely see a different between men and women, I think. Especially at night. So I think there's much more concern about safety for instance about the places that you're willing to travel to. And so I think you would certainly see that. Yeah. I think probably with regard to any demographic if you were to take one demographic and exclude others you would see a different map. >>: So you mentioned that you didn't find quantitative evaluation very convincing. What have you tried? >> Justin Cranshaw: So I spent some time trying to link it to property values. And I think it's a natural thing to try. And you do see some evidence. But it's just not the kind of evidence that I -- that I felt was capturing the spirit of the work, let's say. So you'll find some correlations between like variability of property values within a clusters and outside of a cluster. >>: [inaudible] just internal consistency checks like cross-validation? Because you're doing these parameter checks but they're basically like [inaudible]. >> Justin Cranshaw: No. >>: So they're not automated? >> Justin Cranshaw: No, I haven't done that. It's a natural thing to try. I haven't done it yet. >>: So beyond property values where you try to see some socioeconomic indicators ->> Justin Cranshaw: Not yet. I'm sure you would see some. I think it's a fascinating ->>: Right. >> Justin Cranshaw: But it's not something I've gotten to yet. I'm sure you would see like demographic differences, income differences. >>: Yeah. I mean, I don't know. I mean, census data ->> Justin Cranshaw: Yeah. Absolutely. >>: Go ahead. >>: I [inaudible] I'm wondering if down the road similar to the way that things like [inaudible] you can [inaudible] a restaurant or different attributes if you're planning on like filter out the livehoods. >> Justin Cranshaw: Yeah. >>: [inaudible]. >> Justin Cranshaw: Yeah. I think that's -- I mean, the next step I think is -- so this is kind of an exhibition. I'd like to turn this into a tool, right? And there are various different tools that you could imagine hanging off of a following like this. You could design them for various stakeholders of the city. So, for instance, like real estate development. So you're mentioning a more consumer oriented tool, like it's searching in the city. Yeah. Absolutely. That would be cool. >>: And -- oh. >>: Yeah. >>: Okay. Sure. I have two questions that I think are both about how the technology impacts our mental models of the cities. The first one is whether [inaudible] direct map run comparison between people who use like foursquare, people who don't. >> Justin Cranshaw: Yeah. >>: It's kind ever like what you did, and I just want -- I can see if you used foursquare maybe you sort of shrink distances, like things feel closer than they are. >> Justin Cranshaw: So there was a short paper in CHI last year ->>: Oh, really? >> Justin Cranshaw: -- that did this. It's Henriette Cramer and Frank Bentley from Motorola. >>: [inaudible]. >> Justin Cranshaw: And they did -- actually they replicated the kind of Milgram studies. But they looked at people that used foursquare versus people that didn't. And their conclusions are basically that they have different kind of views of the city. It's an interesting work. >>: I'll take a look. And then the second part of it is you were talking about how neighborhood had sort of shifted and changed. So I'm just wondering about shifting and changing beyond -- so for example this friend of mine over the weekend said that Ballard is the new Belltown, which if you live in Ballard is probably horrifying to you. [laughter]. But his explanation was now that you can go to Ballard and they have valet parking now a sufficient number of places so that people who would go to Belltown are now feeling comfortable to go to Ballard. But anyway, but those are not contiguous. But that's a perfect example of how you can -- you can detect that type of ->> Justin Cranshaw: So I mean I think I'll go back to this. I'll go back to the Seattle example. Only because I finds it so fascinating. If you look at Belltown and you use this -- I think it's this one. And you use this kind of related feature which hides everything else that's not like the top 5 most related of the livehoods, you see way out here -- did I take the wrong one? Yeah. Is this area where there's the wineries. And it's basically highlighting the -- it could be a bias of like a super user, right? But I think it's probably not the case. I think I have enough data that it's not biased by a single super user. I haven't looked at it yet. This was just kind of released. But here you're seeing that like ideally these should be like all continuous contiguous regions, right? I mean, if distance mattered so much, all of the related livehoods would be nearby. Which by and large they are. Except for this one that jumps way over here. And you see that in various other cases of the city too. >>: Would you call that a neighborhood, quote/unquote, including the ->> Justin Cranshaw: I'm not sure I would be comfortable calling any of these neighborhoods. I mean, that's why we used a different name. I mean, so I mean it's representing kind of commercial activities we're not really capturing -- neighborhoods are kind of residential in my mind. So there's some quality of neighborhood to them. But it's not quite a neighborhood. Right? I mean, you'll see that malls sometimes exit at livehoods. But malls aren't neighborhoods, right, they're just -- an airport is the livehood because it's a functional unit of innocence. >>: So, you know, regarding your -- the interviews that you did, the examples that you showed, it seemed like the demographic was mostly like under 30 people who use social media because that's probably how they saw your ad. Were there other demographic people ->> Justin Cranshaw: Of our interview participants? >>: Sorry? >> Justin Cranshaw: You mean the interviewed participants? >>: Yeah. >> Justin Cranshaw: Yeah. So I can go back. I think I have a demography slide. Yeah. So I mean we had people in like 40s -- 40s to 60s actually. >>: Right. All of them essentially are like kind of like [inaudible] in some sense. >> Justin Cranshaw: Yeah. I think by and large with -- yeah, that's true. >>: Did you try showing the maps to someone who doesn't use -- I mean, were these foursquare ->> Justin Cranshaw: Yes. So, I mean, we had I think two participants that were -could be categorized as not tech savvy. And hardly any of them actually were foursquare users. So that's one nice thing about the study. And it would be good to compare. >>: Right. >> Justin Cranshaw: But I think that the density of foursquare users in a random sample is not very high. So it's kind of hard -- a hard study to actually construct. But, yeah, so we did interview some non-tech savvy people and actually have some interesting perspectives. One example is -- I'll go to Pittsburgh. So in Pittsburgh you have this big donut hole right here. There's nothing in the middle. That's not true. This is like -- this is a left-out population from our sample. So this is very low income in the center here. But we interviewed actually residents of this neighborhood. And so one of them said wow that's a visualization of the digital divide. And they're right. That's exactly what it is. >>: Could I ask one more question? Are we -- do we have time? What's your sense of the relationship between -- talking about density at the start, between density and degree of change in neighborhoods? Or is there one? >> Justin Cranshaw: I'd hesitate to answer. But, I mean, my instinct says that densities change faster. Right? That's my instincts. I could be wrong. What's your motivation for asking? What's your thought? >>: I don't know. I'm not sure. In some of the example -- I was just trying to sort of debating in my mind as you were giving some of the examples, I can kind of see both. I could see things that are [inaudible] have more opportunity to change. >> Justin Cranshaw: Yeah. >>: Right? Like once things get sort of settled with density, it's almost harder to change. >> Justin Cranshaw: I could see. >>: Or I could see the other way around too. So I just didn't -- I wasn't sure. >> Justin Cranshaw: It's a function -- yeah, there's different forces, right? Like in a density you typically have larger financial forces at play that have the power to make changes fast if they want to. But you're right. And then there are all these historical considerations in a city that there is a -- I went to a talk where someone showed the layout -- a historical layout of Boston overlaid like on a modern layout. And although the buildings are different, the things that don't change are the actual layouts. Layouts is kind of like an API and, you know, once you set it out, things can change. But the API never changes. >>: Now, it seems -- just one quick comment. It seems to me that, you know, I mean, the main [inaudible] of these changes that certainly has an impact on how neighborhoods change. I mean, I don't know if that's -- [inaudible] big cities in the U.S. But I see in a lot of like big cities in Indian which are expanding very fast because of the growing population that businesses are actually convenient entering out to like other suburbanian areas and growing very fast and redefining the neighborhoods. So that is probably another way. And these changes are really fast, like five years, 10 years neighborhoods change. So [inaudible]. >> Justin Cranshaw: I would say less than areas -- it's rare to see like the construction of a neighborhood. Because the idea is a little less well defined in a less dense area. And it has to do with the variability issue. You don't have the same level of variability in less dense areas. So you will see like new businesses come up, maybe new roads. But you won't see the construction of like South Lake Union. Rebecca? >>: When you're using something besides foursquare as your source? Like what could you use? What are the features ->> Justin Cranshaw: So, I mean, I'm actively in the process of trying to expand my data. So I mean the next things to go for are Instagram is a natural next things. I think it's a different sample. It's also a bias simple. But it's biased in a different way. It might be interesting to compare. But also Twitter, right? I mean, so I -- I don't necessarily like using like -- I find using geotag Tweets is kind of a pain. But what I could do is download all of the Tweets from users that I currently have check-ins for and use the Tweets as sort of a kind of model for their model for the user and have some sort of user to user similarity to incorporate into sort of a -- almost a demography score. Right? So there's different ways of thinking about it. But the point is, I won't make a prediction that foursquare will exist in 10 years. But I think this kind of this kind of data will continues to exist in various forms. >>: It's important that they're actually checking in and choosing to [inaudible] the public that they're there. Or if you had like phone -- and ->> Justin Cranshaw: So, I mean, that's actually another bias that I didn't mention. The fact that these are public Tweets and are public check-ins and not their everyday check-ins is a bias. Because the things that people share publicly on Twitter are different from their kind of everyday activities. >>: [inaudible] foursquare. >> Justin Cranshaw: Yeah. So there have been discussions but never moving beyond discussions. Yeah. >>: [inaudible] doesn't scale very well. Historical atlas. >> Justin Cranshaw: Yeah. >>: Sherlock Holmes, his livehood or Leopold Bloom and gather a few books written about New York and compared the livehoods of their [inaudible]. >> Justin Cranshaw: I like that. It would be fun to see like Sherlock Holmes livehood. >>: Compared to a person living Baker Street 221 doesn't really exist. But in that area. >> Justin Cranshaw: Yeah. I like that. >>: [inaudible] interesting in places like New York or LA comparing tourist maps and sort of the way that those maps are presided by hubs. And if you could remove some of those and compare those path ways compared to regular commercial path ways and see if there's like the citizen livehoods versus the commercial livehoods. >> Justin Cranshaw: Yeah. So even looking at like the difference between tourists and locals is fascinating. There's that fascinating Eric Fisher map. Have you seen that? Has anyone not seen that? I actually have a slide of it. I've just got to find my cursor. Yeah. So this is the -- this is San Francisco. The red are colored as -- this is from Flickr data. The red are tourists. The blue are locals. The yellow are kind of indeterminant. But it's ->>: [inaudible]. >> Justin Cranshaw: It's based on profile information and I guess other data of that person. So where else have they taken photos? >>: [inaudible]. >> Justin Cranshaw: Yeah. He does some fascinating work. But this -- this idea being a tourist, right, you almost have a different persona, right? You're a different person. You have different objectives. You have different -- you're serving some different functional role in the city. So I think that distinction is fascinating. Well, thank you. [applause].