 Hi everyone, I think we're just going to get started and kind of let people settle down. So today we have Levi Wolf presenting on finding the fault lines. Levi is a senior lecturer based in the University of Bristol in the United Kingdom, in case we didn't know. He's also a fellow at the University of Chicago Center for Spatial Data Science, a fellow at the Allen Turing Institute for Data Science in AI in London, and is an affiliate faculty at the University of California Riverside. Levi completed his Ph.D. on gerrymandering in the U.S. at Arizona State University in 2017 with professors Sergio Ray, Luke Anselin, Stuart Fatheringham, and Wendy Cho. Wendy Tam Cho. Currently he works on developing new data scientific and Bayesian statistical methods to improve our understanding of boundaries and bounding an urban social fabric, engaging with political problems, studying the emergence in geographical structure of gerrymandering, and social questions examining barrier movement, settlement, and social ties in neighborhood formation, and economic issues, which is understanding cluster development policy and regional urban economic planning. In the past, Levi has worked in industry as a geographic data scientist at Cardo, my old company, and leads and contributes to many open source geographic data science projects in Python. So with that, let's welcome Levi Wolf. Thanks, Renfe. So today I'm going to be talking a little bit about a broad research agenda that some of my colleagues at the University of California Riverside have put together focusing on boundaries and bounding in urban social fabric. It's pretty important, I think, because the topic of inequality, you know, you may have seen recently Nobel Prize in Economics sent out on inequality and sort of the focus study of this kind of subject, any inequality implies a spatial boundary of some sort. When you have communities and sorting processes, we've known for a long time that this generates social and economic boundaries. And in geography, the study of many different related concepts around boundaries becomes incredibly important. So today I'll talk a little bit about where the current state of the art is on studying boundaries in urban social fabric for economic and social processes. And then I'll talk a little bit about the ways that I aim to break these down in my current work. So fundamentally, it becomes important, I think, to understand an old dichotomy in geographical research. Geography as a discipline is split between the study of what I would describe as place and space. And we'll talk a little bit about how this kind of dichotomy is constructed and how scholars tend to either move between the two or focus on one of those domains as we go forward. Then I'll talk about how boundaries are in both of these notions of geography and how we need to do a better job about thinking about boundaries that move across the two notions rather than only operate within one of the notions. So first off, starting with this old tension in spatial science. When we're thinking about what a boundary is, lots of people describe them in this sort of terms. So here we're talking about Williamsburg, the country when it pushed them. And if you are particularly knowledgeable about Wikipedia, you recognize that this is a description of some sort of process between two neighborhoods that tells you where one neighborhood stops and one neighborhood stops, right? It's a pretty good understanding of a boundary. Then you've got this other kind of definition where we're talking about the built environment. You know that this entity, Greenpoint, is boarded on the southeast by BQE. So in geographical terms, we talk about these things as naturalistic divisions of urban life. And if you've read some of the classic texts on urban planning and urban sociology, you know that this has received quite a bit of attention, how built environments sort of constructs and constrains individual relations. And today I'm not going to talk about anything about this kind of thing. Instead, I'm going to talk a lot about the boundaries between communities. So here is another description that tries to talk about sort of the social fabric of an area. Here I'm now zooming in on Bushwick. And just to your eye, given these kinds of population maps for different ethnicities and races, you can kind of identify edges in the urban social fabric. You can kind of see immediately where a community kind of just jumps out of the image. I don't really have to draw these for you, for your eye to pick up on that pattern across the floor, right? And that's because we're really good at identifying these kinds of patterns of vision. This kind of socially constructed division of urban life is what I'm talking about today. And indeed, it's not just me, it's tons and tons and tons of people. If I stretch this list back to 1925, one of the first foundational texts on urban sociology, PARC, studies this exact phenomenon about sort of community delineation and determination in urban environments. And so there's a lot of different people that have sort of attacked this question of community delineation and identification in a lot of different ways. But recently, I'm very interested in this concept of social frontiers. This has received quite a bit of attention in urban sociology in recent phase because of its link to a very long-standing sociological hypothesis, this contested boundaries hypothesis. So this hypothesis is about communities and their relationships in an urban geography. And it suggests that conflict, somehow operationalized, arises at poorly defined boundaries that separate ethnic and racial groups. Because these boundaries threaten homogenous community life and foster ambiguities about group rank. So thinking about it, the main sort of point here is that ambiguous boundaries threaten homogenous community life. And this somehow drives conflict between communities. But there's kind of a slide of hand going on here that I think is important for you to know. Here we're talking about community. But what's on the tin is about neighborhoods. And this is an ontological distinction. These are not the same things necessarily. Communities are not necessarily a neighborhood. And the neighborhood is not necessarily a community. Neighborhoods have a lot of different definitions from planning units to social constructs about real estate markets. So this hypothesis deserves some really serious conceptual interrogation. It takes the fact that communities are neighborhoods. These neighborhoods are pre-existing. They're territories, distinct areas that limit a given social group. So one territory has one group. And then when this territory is unclear, communities come into conflict. So most of the theory around boundaries in sociology and urban economics are focusing on this kind of definition. A community that has a territory and that community might be a market, that community might be a social community. It's something that sort of is this entity. And when that entity's edges become unclear, it somehow emerges into conflict. So this could become price competition, where people's margins are driven down for selling of houses or renting short term properties. Or it could be, as in the case of the article I'm talking about here, 311 complaints, complaints about noise and vandalism. The point is communities have a territory. We know that a priori. And then when that territory becomes unclear, communities come into conflict. This is how most of the literature on boundaries works today. But the problem is this is a very place-based definition of how boundaries work. In geography, we have two large conceptual areas of research. We have research on space and we have research on place. Some people try and blend these two things. This is a recent paper. I really don't think I will ever find myself saying that word just because it doesn't really roll up the tongue. But that's all right. It's a reasonable stab at an attempt to blend the two. With that said, space and place are usually defined in this kind of way. So the study of space in geography is the study about geographic systems over which objects are related. So this comes down to things like studies of networks, studies of relational systems, studies of that kind of inter-worked network fashion. And then place study tends to be discussed as sort of trying to identify regions or neighborhoods, looking at how those things emerge, how they're constructed, either sociologically or economically. It might be about trying to derive particular typologies. If you've ever heard of the study of the marketing term geodemographics, where you're trying to segment, you know, these are aspiring young people. This is, you know, maybe like an old person community with affluence, retirement wealth. They segment society in all these different places and try and say, well, this is where this type of person lives. This is where this type of person lives. That's place-based research. Space research, on the other hand, is characterized with this network relational thinking. There's a lot of recent scholarship in this place-based identity. Plus, all of that stuff I was talking about Boundaries before is a place-based project. In fact, geography throughout its history has gone through sort of iterations of whether or not we focus on space generally as a domain or place generally as a domain. And if you go through a sort of history of geographic thought, you can see major thinkers painting back and forth between these two perspectives on sort of geographical resistance. Right now, I would say we're at sort of the apex kind of going over a hump of place-based research. And these are the kind of questions that you focus on in place-based research. How or why do these places emerge? Or how are they socially constructed? What are their properties? How can we tell places apart? So if you've ever heard of ideas about clustering or urban regionalization, that's all place-based research. Spatial research in this sense focuses on how things interact. And those entities are not usually defined as sort of emergent or constructive things. So between these two sort of kinds of questions, you can identify whether or not some of your research is place-based or place-based. And as I said before, most of the work that's done on boundaries is focused on place pre-existing or constructive. So this is a selection of a couple of the recent papers on place-based boundaries. And what I'd like to focus on is a methodology that they call Wombling. Wombling is a statistical technique developed in I think 1950s by a person with the unfortunate last name of Wumble, which aims to try and identify the sort of distinctions between different places based on an a priori theory about how places are related to one another. And all these things actually happen to use a multi-level model with a generalized linear regression. So I'm not going to go particularly strongly into these methodologies, but basically this paper up here finds boundaries between pre-existing known neighborhood divisions based on large differences in predictions of a multi-level model of crime in an urban place in the UK. And then this paper looks for sort of urban economic boundaries between real estate markets and liberal. And it tries to identify whether there's discontinuities and pricing, where if I'm on one side of the line, I might be expected to pay tons more money for the same property or effectively the same property. Just a marketing thing, just how things are branded and labeled and sold. But this is also an important part of economic community delineation. And then this paper here is an interesting paper that sort of looks again at the predictions of crime based on sort of theories of relevant ecological or personal factors. And there's a ton more literature on this in epidemiology, which I'm not going to focus on today, because that tends to be a little more cut and dry. But here, this is the application of this epidemiological technique to urban sociologists. So when you do a wobbling study, you get something that looks like this. This is Chicago. And what this, I think this is from the GUI paper. So what's done here is focusing on one particular ethnic or racial group, we come up with a predictive model of the relative excess share of crime based on various ecological determinants. And we get this sort of value of how likely an area is on a boundary or not between two communities. Okay, and we get this is this is the kind of end result. This is what we're trying to sort of generate. And there's a problem with this kind of work because when you try and study community conflict and the contested boundaries hypothesis, you don't really define precisely conflict over what and between whom. All we're doing is saying that the outcome of conflict is sort of crime or specific instances of crime in 311 data, or its price disparities that reflects some kind of economic competition. Our definition of what the actual conflict is is unclear. And we don't justify our input data, we don't modify, we just say, well, it's these generic ethnic racial economic inputs that give us these boundaries. So there's not a really good theory linking why our output variables are related to our input variables and how the boundaries emerge from that. Further, axes of social existence are not independent. Places emerge because they're distinctive and internally similar. So it becomes incredibly impossible to validate any statistical assumptions of these techniques. And I'll show you that in a second. The second thing is, when we study boundaries currently in boundary analysis, we come up with things that are symmetric and reversible. We come up with this idea that if I'm on a boundary, then, you know, it's sort of equally transversible either way. People on one side of the boundary have equal access to that boundary as people on the boundary. So there's no notion of the fact that some groups are more privileged than others in urban social spaces, and may find it easier or harder to own space and territory. Further, and I think most concerningly, all of these approaches that I've talked about, all of these place-based longland approaches, they assume the existence of places and the scales at which social processes operate before they ever try and estimate boundaries. And I'll show you in a second how we can come up with better identifications of boundaries in urban environments without ever referring to specific known places. So first off, all of these studies that look at the boundaries between neighborhoods and try and identify whether or not they're strong or weak or related to crime or not, they assume that these places are distinctive, right? If I tell you that Clinton Hill is my neighborhood of study, you know where that is. You know what it's like. It has a particular demographic profile. It inhabits a location in the city. And you can probably tell me when you're in it and when you're not, right? That's a place-based understanding. But we can sort of think about how these community areas balance between being distinctive and being geographically compact. In the study of places, you get sort of the analysis of individual neighborhoods, which have geographically contiguous territory. And then you might also get studies of people places, sort of these market segments of upwardly mobile, of retired rich people that I was talking about before. And place-based research doesn't really distinguish between those two things in common part. So to illustrate this, I'm going to talk about a little bit of work I've done on spatial machine learning. That looks at manifold learning. Manifold learning aims to take a bunch of different information about social and economic structure and compress it down into just a couple of different variables. And those relationships can be nonlinear. They can be really confusing, really difficult to understand. So if you've heard of like principle components analysis, manifold learning is kind of like a nonlinear principle components analysis. But you can do some interesting things. So here in this paper that I'm linking, I'm analyzing how we can understand urban social and geographic boundaries in really high dimensional data. The census, which if you were at the previous talk, it measures a lot of different things with a lot of uncertainty. There's a lot of information. So how can we look for these urban social boundaries in a way that doesn't essentialize individual places? And what I'm using here in this paper is a technique called ISO. I'm using a spatial elaboration of it. But the idea is that you you're looking for sort of in a multi dimensional space, you're trying to move along a nonlinear relationship. And without knowing the structure of that multi dimensional space without knowing the sort of theoretical high dimensional relationship between race and income and ethnicity and religion and age, I need to come up with a way to try and understand that high dimensional structure. So what the ISO map algorithm does is it says, well, let's hop between different similar kinds of observations and use that as a model of a high dimensional distance. Yeah. Mm hmm. Yeah. This is like the non, this is a vector. Yes. So this is like if we took all of the information we had in the census and visualized it as a point clock. Typical dimension reduction would look at this dashed line as a measure of distance between two points. Here we're talking about two census blocks. But what we need to recognize is that that's kind of a shortcut that moves from one part of this nonlinear relationship to another part of it without ever considering the fact that there's no real jump there. Instead, the correct technique would walk along this relationship until you get to that point. It's not really important to get the specifics of this, but that's how these kinds of nonlinear machine learning techniques work. I think what'll make it more clear is looking at the visualization. So if you start with this learning algorithm tuned to be hyper local, you basically just get the geography of Brooklyn back. So here I'm looking at four covariates about urban social structure. And I'm sort of doing some technique to it, but basically I'm recovering the geographical positions of areas in Brooklyn. So as you see here, there's sort of just a direct map in between these things. It's stretching it, it's rotating it a little bit, but in general we're putting in the positions of census blocks and we're getting back the positions of census blocks. But as you allow the algorithm to model distance, it kind of warps the surface moving blocks that are similar to one another in geography, in attribute terms, moving them close geographically. So you can kind of think about this as working the geography so that you're moving from something that looks like Brooklyn to something that looks very, very different. You're actively taking all of these geographical observations and moving them next to one another when they're similar. And if they're dissimilar, you're trying to push them apart. So that's how it kind of works. Does that make sense? Yeah, that is a map. If I wanted to, I could draw all the census blocks and show you how they're all rearranging. It gets kind of dirty. In this diagram, let's say yellow places have a higher percentage black population. And as we go from focusing on geographic structure here to focusing on the data structure, places that are yellow should be pulled towards one another. But this is operating in eight dimensions. So it's trying to pull places that are all similar on all of these axes closer to one another and push dissimilar areas away from one. When you do this, regardless of sort of how you tune it, and I've taken it out to a lot more than eight dimensions, when we ignore the spatial relationships in places, you kind of get similar geographies after you run it through this sort of black box we project. So when you think about a geographical entity in real space that has a distinct profile of distinct demographic profile, if I run it through a manifold learner, that thing tends to cluster in the same space. Now that takes eight dimensions of our data and that has a geographical location and in its kind of data location, all of those points in the neighborhood are still close to one. And that happens for nearly all of the neighborhoods in Brooklyn. I can run it through them all, but the point is, things that are near one another tend to be similar to one another. Yeah. Right. What is it telling us that's different? Well, the study of boundaries tends to take one of these things as a Y and put the rest of them as an X in a predictive model. But that's improper if they all are just sort of varying together with each other. There's no real reason to expect that crime rates are going to behave anyway independently of economic status or systematic disadvantage. The point is the current literature assumes that they're separable and they're not. So that's why this is important. Here what I'm showing is, in the data geometry, we have effectively the same relationships as geography. The data reproduces geography and geography reproduces the data. We can't assume that these things are separate and try and study boundaries in the data sense without thinking about the geographical sense. And that's exactly what the current literature does. It doesn't acknowledge that the effective dimension of the data is really small. So we know that all of these things are intersectional. Income, race, ethnicity, religion, age, they're all really tightly bound up. And currently the literature will set one of those things as an output and say everything else causes it. And then try and look at our prediction errors in that model. And you can't do that. Geography both expresses and embeds social structure. So when you separate those things out in a model of boundary, it doesn't work. Does that answer your question? The literature assumes that they're sort of separable things, but they're not. So in the boundary literature people will say, well I'm going to search for boundaries in the crime landscape, or I'm going to search for boundaries in the health landscape, looking at life expectancies. But all of these things in both their geography and in the data sets occupy the same places. So a boundary here is just going to recover one of these boundaries. There's no effective enlightenment that happens by studying this data geometry. And we've known this for a long time and adding nonlinear complicated effects doesn't change the fact that if I'm read over here and I'm clustered, I'm probably going to be clustered in the geographic space as well. In geography we think that this is the first law of geography. Near things are more related than distant things. It doesn't matter if I throw it into a fancy nonlinear machine learner or not. Near things are still more related in high-dimensional space no matter what you do. Yeah. So this work does not necessarily pull us away but it illustrates precisely that that Hark model of all of these things being functionally classified together. It's not a data, it's not a technique problem, right? So those studies originally use things like principal components or factor analysis, right? Factorial ecology use something that is a linear version of this. The problem is not methodological and that's what I'm trying to show here. Just assuming that you can use a fancier machine learning algorithm doesn't matter. So the next bit is about sort of how these models of difference are symmetric and reversible, right? If I'm classified as one community and then another group is classified in another community then they're sort of different classes. They're not sort of transversible in either way. They're symmetric and reversible. I'm in one class or I'm in another class. This area is owned by one group or it's owned by another group. And what you can show is that this probably is not the case. So this is a paper that just looks at sort of how these boundaries are analyzed and expressed, again in sort of high-dimensional data. So to do that, I want to talk a little bit about this statistic called a silhouette statistic. It's used in clustering and data science. You tend to have observations like a census block and some graph. Here you can just think of in all of Brooklyn. And we say that it's assigned to some place C and not assigned to some other place K. So you can think about this as a neighborhood has a bunch of different street blocks. Street blocks are I and the neighborhood is either C or any other neighborhood K where those street blocks aren't assigned. Then to define the actual silhouette statistic you come up with something like this. And to break it down, this is the dissimilarity between a block I and its place C. So this is how similar a block is to its community. And this is the dissimilarity between that block and some other K, some other group, where that block is the most similar to it. In the literature, the original suggestor of this technique calls this your second best choice. So if I'm a block in Clinton Hill, this is going to be the next neighborhood that is most similar to that block. This top part is positive when I is more like its current cluster than its second choice. And this is just a normalizing factor that makes sure it falls kind of like a correlation coefficient between negatively. All together this statistic tells us a chronic gap between a block's current place and its second most fit or second best alternative. And this is used all the time in data science to kind of analyze goodness a fit of clusters. It also can be adapted to give us a better understanding of how boundaries can be directional and not symmetrical. Looking at a geography here, where we're sort of looking at the structure of this goodness a fit statistic, we get a bunch of neighborhoods here that are Zillow neighborhoods built from online housing markets. And then we can visualize the most similar alternative assignment for each census block. So this is like if I could pick this up and stick it in another neighborhood, which neighborhood would I put it. And there's a bit of a cartographic lie going on here because I don't have enough colors to represent both of these things effectively. So this isn't necessarily a helpful diagram by itself, but what it gives you a good understanding of is when it comes to the actual silhouette statistic. This is the distribution of those goodness a fits on the real empirical data for demographic information in Brooklyn. And then these are the pre-existing neighborhood boundaries that are drawn by these housing market circles. And with respect to their neighborhood, blue observations are very dissimilar and orange observations are similar. What this means is that most blocks could be better fit into another place. They're more similar to somewhere else than they are to where they are. And you might have heard about sort of ideas about designing for diversity. And you might have sort of heard about the fact that sometimes we actually don't want to think about territory in places as homogenous. This is exactly what this shows. Most of the blocks in Brooklyn are more dissimilar to their surroundings than they are under sort of an optimal clustering model. This says that nearly everything is poorly fit to its community. So that illustrates right here that this whole perspective that we're looking for communities that are occupied by a single group is severely flawed in the study of boundaries. Moving forward we can make this directional. Taking the same statistic all we have to do is swap out what cluster we consider as an alternative. So if before we're looking at how similar a place is to trying to find its best other place you could send it to, this one says okay if I'm a place, if I'm a block and I'm sitting right next to another neighborhood, how similar am I from where I am to where that other neighborhood is right across the street. When we focus on that we're actually sub-setting all of these census blocks to only look at blocks that are on the boundaries between neighborhoods okay and you get something like this when you look at it overall but that's kind of hard to interpret so I'm going to zoom in to a couple of different examples. This is the first one and earlier we were talking about that boundary between Bushwick and Bed-Stuy and how apparent it was to our eyes. When we think about boundaries as directed objects we can see that as well. This statistic when you compute it and look at how similar these boundary observations are if I were to flip them right across the street we see that blocks are more similar to their own community than they are to the one right across the street and likewise if we go on the Bushwick side we see that blocks that are just right on the other side of that boundary are more similar to their group than they are to the one that's right across the street. So this is a sort of mechanical characterization exactly what we just were able to pick out their eyes but this operates in much higher dimensional data. So one thing that we can test is if this boundary is a sharp boundary it shouldn't lead to conflict but focusing on another area it's entirely possible for it to be a symmetric. So here in one neighborhood observations are much more similar to the one across the street than they are to their own but then the ones in Carroll Gardens are much more similar to their own group than the ones on the other side. So you can think of this as this kind of leans one neighborhood into the other. This boundary is a symmetric things on one side of this place are sort of leaning into the ones on the other. You can also have two that are negative. So boundaries are not necessarily just hard discrete objects they can be direct they can lean. This is an asymmetric boundary. In the literature there's no distinction between this and the other. So the very last bit of this talk I think I'll skip over some of the more heavy discussion of similar methodology here but this I think is the most critical problem. All of the current studies of boundaries and like the one I just showed you now assumes that neighborhoods and places are stacked. They're known in advance and conflict operates at a single geographical scale. So we know that urban morphology the structure of communities is not single scale. Urban systems have been well characterized as these things that are called fractal multi-scale systems. So from this quote in this Michael Batty who's a very famous thinker on city science and complexity. We see that city morphology is reflected in the hierarchy of different sub-centers or clusters right. These different places around the city you can come up with sort of broad regions downtown and then you can also think of neighborhoods that operate inside of those regions. And our current study of neighborhoods doesn't really acknowledge this in a coherent way. So because our social relations in the city are embedded within this morphology and we also sort of enforce and adjust that morphology. We have to come up with methods that don't just assume our places that we're studying. We have to come up with methods that don't just reproduce the existing geographies that we've given them. And most of the time that's exactly what happens in this boundary literature. So here anyone who's familiar with Chicago can recognize that if we just send it all of the stuff that we know about Chicago we're going to get a traditional segmentation of Chicago that we recognize. There's nothing new here. We're just proving that we can identify where the south side is based on different information using a different methodology. There's nothing sort of fundamentally intriguing here about what gets interesting is when you think about how this changes on a higher sort of varying state. So here I'm not going to go into the specifics of sort of the information theoretic argument I'm using here. Suffice it to say I'm working on sort of definitions of entropy, studying high-dimensional relationships between different kinds of demographical factors. So this is the entropy of census unit I. It's some function of the population structure in that area. I'm just going to skip over this stuff until we get to here. So what I'm doing is I'm looking at some measure of population heterogeneity. How sort of different is an area of population. And I'm doing that using this idea of an ego hood. An ego hood is kind of the area around an observation. And you can think of D with delta as all of the other blocks that fall within a distance delta of a single neighbor. And when you increase that D you start to see different scales of urban boundaries. Different scales at which blocks are dissimilar. So overall this is a sort of a population weighted ego hood average. I'm not really too concerned about that. The point is we're looking at how different each block is from some surroundings. And we're allowing that definition of what surroundings is to there. We're allowing it to be really local all the way out to be nearly the entire state. And when we do that we get something that looks like this. We're allowing the definition of what the boundary is looking at. The boundary between what distant places. We're allowing that to increase as the animation plays through. And sure at some scale we can identify this boundary. But that's only one scale. That's only one geographical scale at which this boundary geography is represented here. At many many many other scales of analysis. At many other scales of analysis. That doesn't hold. If you re-normalize the colors on this group or you allow the scale to continue going in a regional sense. You can see many other structures that are apparent in where urban social boundaries are. Even when you focus here you can see that there's an urban boundary that differentiates downtown from out of town. As it grows larger and larger. So when you think about this. As there's not just one scale of analysis. It's not just the neighborhood. It's also the region. It's also larger than the city. It's the city region. Boundary study has to be multi-scale. We have to think about the fact that at any one scale of analysis. Boundaries can exist between places. And the current work doesn't do that well. Sure. So I'm suggesting A that it exists. That there's not just one scale of analysis in the city that's important for trying to understand social boundaries. There's a lot of different scales. This whole range of social outcomes matters. And if we just focus on one. Then we're only going to reproduce these existing scales of analysis right. Neighborhoods. When you assume them are priori and study them you get something that looks like this. A static picture of how the urban social landscape is divided. But when you allow those things to consider whole ranges of individual interaction. People move when they go to work. They interact with other communities as they go throughout their day. This becomes much different. So you can you can think about this in terms of like how long does an individual residing in one of these areas commute to work. How often do they pass through communities. That is a multi-scale analysis. That means that we can't just consider the inherited neighborhood. We have to think about it as a multi-scale phenomenon. And the best part about it is that we can recover this map just by picking out one of those scales. And nearly all of the studies of a city do this. You can use this kind of multi-scale thought process. And come up with all of this replicated entire studies worth of assumptions. Using just a very simple scan technique. So the point here is that we can recover all of these very complicated theories about how places relate to one another. Without assuming that they all operated a single scale. And we can come up with higher level and lower level operationalizations of that scale. So it could be the case that you need to find boundaries that are sub-neighbor. It could be the case that you need to find boundaries that are much larger than that to differentiate regional labor markets inside of the city. The point is this view is fixed and static. And we have to be aware of it. So if my encouragement would be that something like this, an analysis that allows these kinds of visualizations to change, is where we need to take boundary study more generally. So with that I think I'll wrap up. The point of this talk was that when we think about boundaries in the city, we think about the study of boundaries in sociology, it becomes really important that we think about most of the places that we're studying. We're assuming them a priori. And we need to instead recognize the fact that a lot of these relationships are complex and multi-dimensional. But oftentimes can't be separated into a clear, live value and X value in a multi-level regression. We need to understand the fact that boundaries can be symmetric or boundaries can be asymmetric and not necessarily have to be reversible. Places can be directionally dissimilar. Especially when we're talking about boundaries, we need to think about the fact that some places are demographically similar to nearby things in one way and not the other. And finally, we need to recognize the fact that cities are very complex, multi-scale phenomena. And when we just assume that the neighborhood is a static fixed thing at a single geographical scale, we are only going to reproduce the thing that we set out and informed our model to begin with. No matter what fancy machine learning or data scientific technique you use, you need to make sure that your thinking is the thing that's changing, not necessarily the technique. So with that, that was my sort of presentation of this work that I've done in geographical machine learning on estimating boundaries and bounding urban social phenomena. Thank you very much. Yeah, we actually see that quite a bit. This idea that, you know, a boundary people can think of as like a large front that cuts across the city. But oftentimes that's not actually a reasonable way to think about it, right? Due to the way that transit links affect the way that objects are related to one another, oftentimes we actually see these kinds of islands or two areas that should be functionally linked that aren't actually geographically next to one another. That's kind of at that multi-scale idea. If we adapted that instead of using sort of spatial distance to something like transit distance or commute travel affinities, you do see those kinds of things emerge. So here I'm just trying to make the point that we have to do that in boundary study and do not currently, but those things are incredibly complicated. Yeah, there's a lot on like the agent-based modeling and trying to segment a city based on potential commute paths and things like that, as well as some spatial social network analysis that does that kind of stuff as well. Yeah, sure. Relationality, mass, yeah, I mean understanding it's on to places, boundaries, what is left? Is there anything useful to place? That's a good question. The first, sure, I'd say the first answer to your question comes from trying to use techniques that don't make a one-to-one correspondence between community and geography. So techniques like that include something like a geographically weighted regression. Some of these sort of scale agnostic statistical techniques where you're thinking about changing the scale of analysis as a parameter to your function, those kinds of things are incredibly important. And I think there's been no sort of equivalent attempt at doing a wobbling analysis on a model like that. No one's looked at it, no one's done it because they've assumed that the places that are conflicting are discrete known entities that clack across an urban surface. So something like that becomes important from a plywood. The second part is what is left of place if we do this? That's a really good question. I think there's a lot of ways that place study is really valuable, but I also think it's intrinsically essentializing. And I'm very uncomfortable with assuming that these places are, I mean they have meaning, right? That's why we use them. They encode something about the urban social landscape. But I also think that as sort of like someone interested in machine learning and data science, those might emerge from the data if we orient the study correctly, but we need to be really careful about feeding that to the learning algorithm and then getting it back. And I think a lot of the boundary study literature is doing that. It's feeding it the information and then getting back exactly what it knows. So with that said, if we do this kind of boundary-seeking activity, we might learn that there are many different scales at which the city is divided, and some of those that we focused on in the past maybe aren't the most salient ones for individuals. I'm not quite sure because the work hasn't been done yet. So lots of these things were preference or sort of discussions that haven't been published yet. So this NSF grant is looking at how neighborhood sort of changed demographically and then also shift in their boundaries. So things like that become important. You need to stop thinking about sort of these geographical entities as fixed territorially in order to do that. And that's really hard methodologically speaking. So thanks for your question. What could be possible? So neighborhood-oriented planning is one thing that really serves to strengthen this kind of an analytical lock. I'm not actually sure that this is a problem from sort of the endogenous, like I don't think individuals thinking about their neighborhood as a community is a problem. What I think is problematic is when academics use a neighborhood geography as shorthand for the community. So a lot of this study is placed based in the sense that it's using places as shorthands for community. And that's where I think a lot of the more interesting things are like, I mean I'm sure people have seen maybe the Bostonography Projects where you're asking people to sort of draw where they think a particular neighborhood is and you come up with these consensus measures. The different ways that individuals think about their own community and the individual sort of takes different paths to work and engages with different businesses and things like that. Academics need to use those new forms of data rather than just relying on cadastral census data to say, this is the place that we're studying and this is the place geography. We might need to come up with new place geographies. So that would be my recommendation. Yeah. So it's closer to one part of the, one side of the boundary, but they also want, it's definitely the other side. So it's definitely the amount of adage rendered to work. So I feel like there's something distinct about boundary stuff, like it's not just a slight feel. I feel like it's very, it's a good answer. The process of drawing a boundary can create those kinds of, what I would have called in like an econometric sense an externality, right? That people are deriving benefit from that. I saw that really strongly in my dissertation work on gerrymandering. Where the people drawing boundaries are sometimes splitting communities in such a way that that community that sits on the boundary is actually more internally similar than it is to places that are, and you can see that using those kinds of directional approaches that I talked about. But at the same time, like when it begins an agent decision where like that's something you elect to do versus where that's something that you've been forced to by a line drawer, that becomes sort of different subjectively. But from a statistical perspective, yeah, that's why I think this kind of study is important. Instead of assuming that it's homogeneous and unidirectional, we need to allow for the fact that sometimes these place geographies are not the real meaningful scales of analysis. Yeah, yeah, that's a real interesting question. I have two projects that do that. So one of them is my project of the Allentarian Institute looking at new cities in Africa and how they grow. And in particular, how their sort of sectoral structure grows. A lot of that stuff is really interesting because territorial governments are usually not updated. So like a city boundary in the US, you might have like various kinds of incorporation agglomerations that happen. But in a lot of the really fast-grant cities, our study site is in Nigeria that doesn't actually happen. So you get there's unlisted slums and things like that that just grow in formal settlement areas that have no effective central governor. So we've got some stuff looking at sort of how that urban structure sort of grows and how we can detect that without having to rely on census information. And we've got some sort of, can you draw your neighborhood or can you draw your route to work kind of stuff going on there? But the one that I'm actually more interested in is in Larkana, Pakistan. Because in Pakistan, they have a really strong census geography that's inherited from sort of British planning traditions and it has been maintained in the way that the sort of settlements are listed and administered. So looking at the ways in which people in Larkana, Pakistan engage with that sort of imposed definition of space has been really interesting. We've got some fieldwork that's gonna be going on in the spring that's asking people to validate these maps that come from administrative sources about whether or not sort of, do you recognize this to be your neighborhood? Do you recognize this to be your neighborhood? Does your sort of walk to work, take you through a bunch of these areas or do you tend to stay in the area where they've sort of said, this is a reasonable labor market kind of thing? So I'm definitely interested in those sort of alternative paths to urban structure. But with some of the sort of data science stuff it becomes necessary to study places with availability. So I am cognizant of that but it's a little harder usually requires a different set of skills. Yeah, that's a good question. Yeah, that's a good question. I did have a project getting started in the UK asking people about why they made the decision to relocate in large longitudinal studies and sort of trying to get at this sense of if they're trying to locate in particular communities based on their like a height factor or things like that. But that's a pretty difficult thing to get at because again, when you think about the fact that a lot of this longitudinal census data has only been available in a clean format for about four or five years, lots of that kind of even further, deeper longitudinal study is very difficult to do on this scale. So I haven't done anything yet. I've submitted some proposals but we'll see what happens when it can make. Yeah, I think I haven't really thought a lot about that. I do think that one thing that becomes important to think about is sort of spreading the pot more widely. I mean like I know in my local context in Bristol, there's a lot of urban development projects that are really tightly focused on single wards and that tends to generate these kinds of externalities of displacement and sort of government led gentrification. And so the analysis of where those boundaries are in the socio-demographic context means that at least in my context, the funding needs to be spread a little bit more broadly and that's so intensively focused on one specific area because that area, that government unit doesn't necessarily match the actual disadvantaged communities. But that becomes difficult, right? Because then you're arguing that the classic administrative lens needs to be changed. But in my experience, that's what it comes down to. We have to do a better job as academics of questioning that when we're assisting policy. Sometimes that means we need to be comfortable drawing better lines. So what do you think about that physical boundaries that are growing and how we can solve that because the little kids that are growing there, this is the reality that they know and what do you think? I admit that I haven't done quite as much on the sort of imposition policy perspective. I know I have some colleagues back in the UK that studied the peace walls built in Belfast and how the imposition of physical boundaries can affect community relations in the way that two really strongly different communities now can identify have less conflict because they interact less. I'm not so much on that particular angle of this. I'm much more on the sort of standing up at the edge of this literature, trying to say we need to think about these things in a complex, multi-relational framework. But yeah, I don't know. It's a fascinating process. I admit that I only know a little bit about sort of the sectarian divisions of Northern Ireland and Scotland. But in other contexts, I mean, it becomes way important to be aware of how these things are both, as I admitted in the talk, sort of enacted by the city in addition to experiences. So that's not something I currently focus a lot on. Thanks a lot, folks.